common : streamline the formatting of help (#4890)

* common : streamline the formatting of help

- Separate alternative parameters by a comma

- Do not indent `--version` differently

* Update common/common.cpp

---------

Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

.github/workflows/nix-flake-update.yml

@@ -19,4 +19,4 @@ jobs:
           pr-labels: |
             nix
           pr-reviewers: philiptaron,SomeoneSerge
-          token: ${{ secrets.GITHUB_TOKEN }}
+          token: ${{ secrets.FLAKE_TOKEN }}

.gitignore

@@ -51,6 +51,7 @@ models-mnt
 /lookup
 /main
 /metal
+/passkey
 /perplexity
 /q8dot
 /quantize

Makefile

@@ -1,8 +1,8 @@
 # Define the default target now so that it is always the first target
 BUILD_TARGETS = \
-	main quantize quantize-stats perplexity embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
+	main quantize quantize-stats perplexity imatrix embedding vdot q8dot train-text-from-scratch convert-llama2c-to-ggml \
 	simple batched batched-bench save-load-state server gguf llama-bench libllava.a llava-cli baby-llama beam-search  \
-	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup tests/test-c.o
+	speculative infill tokenize benchmark-matmult parallel finetune export-lora lookahead lookup passkey tests/test-c.o
 
 # Binaries only useful for tests
 TEST_TARGETS = \
@@ -614,6 +614,9 @@ quantize-stats: examples/quantize-stats/quantize-stats.cpp    build-info.o ggml.
 perplexity: examples/perplexity/perplexity.cpp                ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+imatrix: examples/imatrix/imatrix.cpp                         ggml.o llama.o $(COMMON_DEPS) $(OBJS)
+	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
+
 embedding: examples/embedding/embedding.cpp                   ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
@@ -665,6 +668,9 @@ lookahead: examples/lookahead/lookahead.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS
 lookup: examples/lookup/lookup.cpp ggml.o llama.o $(COMMON_DEPS) $(OBJS)
 	$(CXX) $(CXXFLAGS) $(filter-out %.h,$^) -o $@ $(LDFLAGS)
 
+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)

Package.swift

@@ -14,14 +14,14 @@ let package = Package(
         .library(name: "llama", targets: ["llama"]),
     ],
     dependencies: [
-        .package(url: "https://github.com/ggerganov/ggml.git", .branch("master"))
+        .package(url: "https://github.com/ggerganov/ggml.git", .branch("release"))
     ],
     targets: [
         .target(
             name: "llama",
             dependencies: ["ggml"],
             path: ".",
-            exclude: [],
+            exclude: ["ggml-metal.metal"],
             sources: [
                 "llama.cpp",
             ],

README.md

@@ -10,6 +10,7 @@ Inference of [LLaMA](https://arxiv.org/abs/2302.13971) model in pure C/C++
 
 ### Hot topics
 
+- New SOTA quantized models, including pure 2-bits: https://huggingface.co/ikawrakow
 - Collecting Apple Silicon performance stats:
   - M-series: https://github.com/ggerganov/llama.cpp/discussions/4167
   - A-series: https://github.com/ggerganov/llama.cpp/discussions/4508
@@ -136,6 +137,7 @@ as the main playground for developing new features for the [ggml](https://github
 - [semperai/amica](https://github.com/semperai/amica)
 - [psugihara/FreeChat](https://github.com/psugihara/FreeChat)
 - [ptsochantaris/emeltal](https://github.com/ptsochantaris/emeltal)
+- [iohub/collama](https://github.com/iohub/coLLaMA)
 
 ---
 

common/common.cpp

@@ -220,6 +220,20 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.n_ctx = std::stoi(argv[i]);
+        } else if (arg == "--grp-attn-n" || arg == "-gan") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+
+            params.grp_attn_n = std::stoi(argv[i]);
+        } else if (arg == "--grp-attn-w" || arg == "-gaw") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+
+            params.grp_attn_w = std::stoi(argv[i]);
         } else if (arg == "--rope-freq-base") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -616,6 +630,12 @@ bool gpt_params_parse_ex(int argc, char ** argv, gpt_params & params) {
                 break;
             }
             params.ppl_stride = std::stoi(argv[i]);
+        } else if (arg == "-ptc" || arg == "--print-token-count") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            params.n_print = std::stoi(argv[i]);
         } else if (arg == "--ppl-output-type") {
             if (++i >= argc) {
                 invalid_param = true;
@@ -798,7 +818,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("\n");
     printf("options:\n");
     printf("  -h, --help            show this help message and exit\n");
-    printf("      --version         show version and build info\n");
+    printf("  --version             show version and build info\n");
     printf("  -i, --interactive     run in interactive mode\n");
     printf("  --interactive-first   run in interactive mode and wait for input right away\n");
     printf("  -ins, --instruct      run in instruction mode (use with Alpaca models)\n");
@@ -895,7 +915,7 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        number of layers to store in VRAM\n");
     printf("  -ngld N, --n-gpu-layers-draft N\n");
     printf("                        number of layers to store in VRAM for the draft model\n");
-    printf("  -ts SPLIT --tensor-split SPLIT\n");
+    printf("  -ts SPLIT, --tensor-split SPLIT\n");
     printf("                        how to split tensors across multiple GPUs, comma-separated list of proportions, e.g. 3,1\n");
     printf("  -mg i, --main-gpu i   the GPU to use for scratch and small tensors\n");
 #ifdef GGML_USE_CUBLAS
@@ -904,6 +924,10 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("                        Not recommended since this is both slower and uses more VRAM.\n");
 #endif // GGML_USE_CUBLAS
 #endif
+    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");
@@ -926,6 +950,8 @@ void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
     printf("  --override-kv KEY=TYPE:VALUE\n");
     printf("                        advanced option to override model metadata by key. may be specified multiple times.\n");
     printf("                        types: int, float, bool. example: --override-kv tokenizer.ggml.add_bos_token=bool:false\n");
+    printf("  -ptc N, --print-token-count N\n");
+    printf("                        print token count every N tokens (default: %d)\n", params.n_print);
     printf("\n");
 #ifndef LOG_DISABLE_LOGS
     log_print_usage();
@@ -1029,6 +1055,9 @@ struct llama_model_params llama_model_params_from_gpt_params(const gpt_params &
 }
 
 static ggml_type kv_cache_type_from_str(const std::string & s) {
+    if (s == "f32") {
+        return GGML_TYPE_F32;
+    }
     if (s == "f16") {
         return GGML_TYPE_F16;
     }

common/common.h

@@ -62,6 +62,9 @@ struct gpt_params {
     int32_t main_gpu                        = 0;     // the GPU that is used for scratch and small tensors
     float   tensor_split[LLAMA_MAX_DEVICES] = {0};   // how split tensors should be distributed across GPUs
     int32_t n_beams                         = 0;     // if non-zero then use beam search of given width.
+    int32_t grp_attn_n                      = 1;     // group-attention factor
+    int32_t grp_attn_w                      = 512;   // group-attention width
+    int32_t n_print                         = -1;    // print token count every n tokens (-1 = disabled)
     float   rope_freq_base                  = 0.0f;  // RoPE base frequency
     float   rope_freq_scale                 = 0.0f;  // RoPE frequency scaling factor
     float   yarn_ext_factor                 = -1.0f; // YaRN extrapolation mix factor
@@ -240,4 +243,3 @@ void dump_kv_cache_view(const llama_kv_cache_view & view, int row_size = 80);
 
 // Dump the KV cache view showing individual sequences in each cell (long output).
 void dump_kv_cache_view_seqs(const llama_kv_cache_view & view, int row_size = 40);
-

convert-hf-to-gguf.py

@@ -817,10 +817,17 @@ class PersimmonModel(Model):
         hidden_size = self.hparams["hidden_size"]
 
         self.gguf_writer.add_name('persimmon-8b-chat')
+        self.gguf_writer.add_context_length(self.hparams["max_position_embeddings"])
         self.gguf_writer.add_embedding_length(hidden_size)
         self.gguf_writer.add_block_count(block_count)
         self.gguf_writer.add_feed_forward_length(self.hparams["intermediate_size"])
-        self.gguf_writer.add_rope_dimension_count(hidden_size // head_count)
+
+        # NOTE: not sure about this change - why does the model not have a rope dimension count when it is smaller
+        #       than the head size?
+        #       ref: https://github.com/ggerganov/llama.cpp/pull/4889
+        # self.gguf_writer.add_rope_dimension_count(hidden_size // head_count)
+        self.gguf_writer.add_rope_dimension_count(hidden_size // head_count // 2)
+
         self.gguf_writer.add_head_count(head_count)
         self.gguf_writer.add_head_count_kv(head_count_kv)
         self.gguf_writer.add_rope_freq_base(self.hparams["rope_theta"])

examples/CMakeLists.txt

@@ -31,10 +31,12 @@ else()
     add_subdirectory(quantize-stats)
     add_subdirectory(save-load-state)
     add_subdirectory(simple)
+    add_subdirectory(passkey)
     add_subdirectory(speculative)
     add_subdirectory(lookahead)
     add_subdirectory(lookup)
     add_subdirectory(train-text-from-scratch)
+    add_subdirectory(imatrix)
     if (LLAMA_METAL)
         add_subdirectory(metal)
     endif()

examples/batched/batched.cpp

@@ -69,6 +69,7 @@ int main(int argc, char ** argv) {
 
     std::vector<llama_token> tokens_list;
     tokens_list = ::llama_tokenize(model, params.prompt, true);
+
     const int n_kv_req = tokens_list.size() + (n_len - tokens_list.size())*n_parallel;
 
     // initialize the context

examples/imatrix/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET imatrix)
+add_executable(${TARGET} imatrix.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/imatrix/imatrix.cpp

@@ -0,0 +1,380 @@
+#include "common.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <ctime>
+#include <sstream>
+#include <thread>
+#include <mutex>
+#include <vector>
+#include <fstream>
+#include <unordered_map>
+#include <algorithm>
+
+#if defined(_MSC_VER)
+#pragma warning(disable: 4244 4267) // possible loss of data
+#endif
+
+struct Stats {
+    std::vector<float> values;
+    int ncall = 0;
+};
+
+struct StatParams {
+    std::string ofile = "imatrix.dat";
+    int         n_output_frequency = 10;
+    int         verbosity = 1;
+    bool        collect_output_weight = false;
+};
+
+class IMatrixCollector {
+public:
+    IMatrixCollector() = default;
+    void set_parameters(StatParams&& params) { m_params = std::move(params); }
+    void collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1);
+    void save_imatrix() const;
+private:
+    std::unordered_map<std::string, Stats> m_stats;
+    StatParams                             m_params;
+    std::mutex                             m_mutex;
+    int                                    m_last_call = 0;
+};
+
+void IMatrixCollector::collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1) {
+    if (src1->ne[1] < 16 || src1->type != GGML_TYPE_F32) return;
+    if (!(strncmp(src0->name, "blk.", 4) == 0 || (m_params.collect_output_weight && strcmp(src0->name, "output.weight") == 0))) return;
+    std::lock_guard<std::mutex> lock(m_mutex);
+    auto& e = m_stats[src0->name];
+    if (e.values.empty()) {
+        e.values.resize(src1->ne[0], 0);
+    }
+    else if (e.values.size() != (size_t)src1->ne[0]) {
+        fprintf(stderr, "Oops: inconsistent size for %s (%d vs %d)\n", src0->name, (int)e.values.size(), (int)src1->ne[0]);
+        exit(1); //GGML_ASSERT(false);
+    }
+    ++e.ncall;
+    if (m_params.verbosity > 1) {
+        printf("%s[%d]: %s, %d x %d, %d\n",__func__,m_last_call,src0->name,(int)src1->ne[0],(int)src1->ne[1],(int)src1->type);
+    }
+    for (int row = 0; row < (int)src1->ne[1]; ++row) {
+        const float * x = (const float *)src1->data + row * src1->ne[0];
+        for (int j = 0; j < (int)src1->ne[0]; ++j) {
+            e.values[j] += x[j]*x[j];
+        }
+    }
+    if (e.ncall > m_last_call) {
+        m_last_call = e.ncall;
+        if (m_last_call % m_params.n_output_frequency == 0) {
+            save_imatrix();
+        }
+    }
+}
+
+void IMatrixCollector::save_imatrix() const {
+    const char * fname = m_params.ofile.empty() ? "imatrix.dat" : m_params.ofile.c_str();
+    std::ofstream out(fname, std::ios::binary);
+    int n_entries = m_stats.size();
+    out.write((const char*)&n_entries, sizeof(n_entries));
+    for (auto& p : m_stats) {
+        int len = p.first.size();
+        out.write((const char*)&len, sizeof(len));
+        out.write(p.first.c_str(), len);
+        out.write((const char*)&p.second.ncall, sizeof(p.second.ncall));
+        int nval = p.second.values.size();
+        out.write((const char*)&nval, sizeof(nval));
+        if (nval > 0) out.write((const char*)p.second.values.data(), nval*sizeof(float));
+    }
+    if (m_params.verbosity > 0) {
+        fprintf(stderr, "\n%s: stored collected data after %d chunks in %s\n",__func__,m_last_call,fname);
+    }
+}
+
+static IMatrixCollector g_collector;
+
+static void ik_collect_imatrix(const struct ggml_tensor * src0, const struct ggml_tensor * src1) {
+    g_collector.collect_imatrix(src0, src1);
+}
+
+
+struct results_log_softmax {
+    double log_softmax;
+    float  logit;
+    float  prob;
+};
+
+static std::vector<float> softmax(const std::vector<float>& logits) {
+    std::vector<float> probs(logits.size());
+    float max_logit = logits[0];
+    for (float v : logits) {
+        max_logit = std::max(max_logit, v);
+    }
+    double sum_exp = 0.0;
+    for (size_t i = 0; i < logits.size(); i++) {
+        // Subtract the maximum logit value from the current logit value for numerical stability
+        const float logit = logits[i] - max_logit;
+        const float exp_logit = expf(logit);
+        sum_exp += exp_logit;
+        probs[i] = exp_logit;
+    }
+    for (size_t i = 0; i < probs.size(); i++) {
+        probs[i] /= sum_exp;
+    }
+    return probs;
+}
+
+static results_log_softmax log_softmax(int n_vocab, const float * logits, int tok) {
+    float max_logit = logits[0];
+    for (int i = 1; i < n_vocab; ++i) {
+        max_logit = std::max(max_logit, logits[i]);
+    }
+    double sum_exp = 0.0;
+    for (int i = 0; i < n_vocab; ++i) {
+        sum_exp += expf(logits[i] - max_logit);
+    }
+    return {logits[tok] - max_logit - log(sum_exp), logits[tok], expf(logits[tok] - max_logit) / (float) sum_exp};
+}
+
+static void process_logits(
+    int n_vocab, const float * logits, const int * tokens, int n_token, std::vector<std::thread> & workers,
+    double & nll, double & nll2, float * logit_history, float * prob_history
+) {
+    std::mutex mutex;
+    int counter = 0;
+    auto compute = [&mutex, &counter, &nll, &nll2, logit_history, prob_history, n_vocab, logits, tokens, n_token] () {
+        double local_nll  = 0;
+        double local_nll2 = 0;
+        while (true) {
+            std::unique_lock<std::mutex> lock(mutex);
+            int i = counter++;
+            if (i >= n_token) {
+                nll += local_nll; nll2 += local_nll2;
+                break;
+            }
+            lock.unlock();
+            const results_log_softmax results = log_softmax(n_vocab, logits + i*n_vocab, tokens[i+1]);
+            const double v = -results.log_softmax;
+            local_nll += v;
+            local_nll2 += v*v;
+
+            logit_history[i] = results.logit;
+            prob_history[i]  = results.prob;
+        }
+    };
+    for (auto & w : workers) {
+        w = std::thread(compute);
+    }
+    compute();
+    for (auto & w : workers) {
+        w.join();
+    }
+}
+
+static bool compute_imatrix(llama_context * ctx, const gpt_params & params) {
+
+    const bool add_bos = llama_should_add_bos_token(llama_get_model(ctx));
+    const int n_ctx = llama_n_ctx(ctx);
+
+    auto tim1 = std::chrono::high_resolution_clock::now();
+    fprintf(stderr, "%s: tokenizing the input ..\n", __func__);
+
+    std::vector<llama_token> tokens = ::llama_tokenize(ctx, params.prompt, add_bos);
+
+    auto tim2 = std::chrono::high_resolution_clock::now();
+    fprintf(stderr, "%s: tokenization took %g ms\n",__func__,1e-3*std::chrono::duration_cast<std::chrono::microseconds>(tim2-tim1).count());
+
+    if (int(tokens.size()) < 2*n_ctx) {
+        fprintf(stderr, "%s: you need at least %d tokens for a context of %d tokens\n",__func__,2*n_ctx,
+                n_ctx);
+        fprintf(stderr, "%s: the data file you provided tokenizes to only %zu tokens\n",__func__,tokens.size());
+        return false;
+    }
+
+    std::vector<float> logit_history;
+    logit_history.resize(tokens.size());
+
+    std::vector<float> prob_history;
+    prob_history.resize(tokens.size());
+
+    const int n_chunk_max = tokens.size() / n_ctx;
+
+    const int n_chunk = params.n_chunks < 0 ? n_chunk_max : std::min(params.n_chunks, n_chunk_max);
+    const int n_vocab = llama_n_vocab(llama_get_model(ctx));
+    const int n_batch = params.n_batch;
+
+    int count = 0;
+    double nll = 0.0;
+    double nll2 = 0.0;
+
+    fprintf(stderr, "%s: computing over %d chunks with batch_size %d\n", __func__, n_chunk, n_batch);
+
+    std::vector<std::thread> workers(std::thread::hardware_concurrency() - 1);
+
+    for (int i = 0; i < n_chunk; ++i) {
+        const int start =     i * n_ctx;
+        const int end   = start + n_ctx;
+
+        const int num_batches = (n_ctx + n_batch - 1) / n_batch;
+
+        std::vector<float> logits;
+
+        const auto t_start = std::chrono::high_resolution_clock::now();
+
+        // clear the KV cache
+        llama_kv_cache_clear(ctx);
+
+        for (int j = 0; j < num_batches; ++j) {
+            const int batch_start = start + j * n_batch;
+            const int batch_size  = std::min(end - batch_start, n_batch);
+
+            // save original token and restore it after eval
+            const auto token_org = tokens[batch_start];
+
+            // add BOS token for the first batch of each chunk
+            if (add_bos && j == 0) {
+                tokens[batch_start] = llama_token_bos(llama_get_model(ctx));
+            }
+
+            if (llama_decode(ctx, llama_batch_get_one(tokens.data() + batch_start, batch_size, j * n_batch, 0))) {
+                fprintf(stderr, "%s : failed to eval\n", __func__);
+                return false;
+            }
+
+            // restore the original token in case it was set to BOS
+            tokens[batch_start] = token_org;
+
+            const auto * batch_logits = llama_get_logits(ctx);
+            logits.insert(logits.end(), batch_logits, batch_logits + batch_size * n_vocab);
+        }
+
+        const auto t_end = std::chrono::high_resolution_clock::now();
+
+        if (i == 0) {
+            const float t_total = std::chrono::duration<float>(t_end - t_start).count();
+            fprintf(stderr, "%s: %.2f seconds per pass - ETA ", __func__, t_total);
+            int total_seconds = (int)(t_total * n_chunk);
+            if (total_seconds >= 60*60) {
+                fprintf(stderr, "%d hours ", total_seconds / (60*60));
+                total_seconds = total_seconds % (60*60);
+            }
+            fprintf(stderr, "%.2f minutes\n", total_seconds / 60.0);
+        }
+
+        const int first = n_ctx/2;
+        process_logits(n_vocab, logits.data() + first*n_vocab, tokens.data() + start + first, n_ctx - 1 - first,
+                       workers, nll, nll2, logit_history.data() + start + first, prob_history.data() + start + first);
+        count += n_ctx - first - 1;
+
+        printf("[%d]%.4lf,", i + 1, std::exp(nll / count));
+        fflush(stdout);
+    }
+    printf("\n");
+
+    nll2 /= count;
+    nll /= count;
+    const double ppl = exp(nll);
+    nll2 -= nll * nll;
+    if (nll2 > 0) {
+        nll2 = sqrt(nll2/(count-1));
+        printf("Final estimate: PPL = %.4lf +/- %.5lf\n", ppl, nll2*ppl);
+    } else {
+        printf("Unexpected negative standard deviation of log(prob)\n");
+    }
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+
+    StatParams sparams;
+    std::vector<char*> args;
+    args.push_back(argv[0]);
+    int iarg = 1;
+    for (; iarg < argc-1; ++iarg) {
+        std::string arg{argv[iarg]};
+        if (arg == "-o" || arg == "--output-file") {
+            sparams.ofile = argv[++iarg];
+        }
+        else if (arg == "-ofreq" || arg == "--output-frequency") {
+            sparams.n_output_frequency = std::stoi(argv[++iarg]);
+        }
+        else if (arg == "-ow" || arg == "--output-weight") {
+            sparams.collect_output_weight = std::stoi(argv[++iarg]);
+        }
+        else if (arg == "--verbosity") {
+            sparams.verbosity = std::stoi(argv[++iarg]);
+        } else {
+            args.push_back(argv[iarg]);
+        }
+    }
+    if (iarg < argc) {
+        args.push_back(argv[iarg]);
+    }
+
+    gpt_params params;
+    params.n_batch = 512;
+    if (!gpt_params_parse(args.size(), args.data(), params)) {
+        return 1;
+    }
+
+    g_collector.set_parameters(std::move(sparams));
+
+    ggml_set_imatrix_collection(ik_collect_imatrix);
+
+    params.logits_all = true;
+    params.n_batch = std::min(params.n_batch, params.n_ctx);
+
+    print_build_info();
+
+    if (params.seed == LLAMA_DEFAULT_SEED) {
+        params.seed = time(NULL);
+    }
+
+    fprintf(stderr, "%s: seed  = %u\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+    if (params.random_prompt) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    llama_backend_init(params.numa);
+
+    llama_model * model;
+    llama_context * ctx;
+
+    // load the model and apply lora adapter, if any
+    std::tie(model, ctx) = llama_init_from_gpt_params(params);
+    if (model == NULL) {
+        fprintf(stderr, "%s: error: unable to load model\n", __func__);
+        return 1;
+    }
+
+    const int n_ctx_train = llama_n_ctx_train(model);
+    if (params.n_ctx > n_ctx_train) {
+        fprintf(stderr, "%s: warning: model was trained on only %d context tokens (%d specified)\n",
+                __func__, n_ctx_train, params.n_ctx);
+    }
+
+    // print system information
+    {
+        fprintf(stderr, "\n");
+        fprintf(stderr, "%s\n", get_system_info(params).c_str());
+    }
+
+    bool OK = compute_imatrix(ctx, params);
+    if (!OK) {
+        return 1;
+    }
+
+    g_collector.save_imatrix();
+
+    llama_print_timings(ctx);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    return 0;
+}

examples/llama.swiftui/README.md

@@ -1,7 +1,12 @@
-# llama.swiftui
+# llama.cpp/examples/llama.swiftui
 
-Local inference of llama.cpp on an iPhone.
-So far I only tested with starcoder 1B model, but it can most likely handle 7B models as well.
+Local inference of llama.cpp on an iPhone. This is a sample app that can be used as a starting
+point for more advanced projects.
+
+For usage instructions and performance stats, check the following discussion: https://github.com/ggerganov/llama.cpp/discussions/4508
+
+![image](https://github.com/ggerganov/llama.cpp/assets/1991296/2b40284f-8421-47a2-b634-74eece09a299)
+
+Video demonstration:
 
 https://github.com/bachittle/llama.cpp/assets/39804642/e290827a-4edb-4093-9642-2a5e399ec545
-

examples/llava/clip.cpp

@@ -126,24 +126,7 @@ static struct ggml_tensor * get_tensor(struct ggml_context * ctx, const std::str
 }
 
 static std::string get_ftype(int ftype) {
-    switch (ftype) {
-    case 0:
-        return "f32";
-    case 1:
-        return "f16";
-    case 2:
-        return "q4_0";
-    case 3:
-        return "q4_1";
-    case 6:
-        return "q5_0";
-    case 7:
-        return "q5_1";
-    case 8:
-        return "q8_0";
-    default:
-        throw std::runtime_error(format("%s: Unrecognized file type: %d\n", __func__, ftype));
-    }
+    return ggml_type_name(static_cast<ggml_type>(ftype));
 }
 
 //
@@ -533,6 +516,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     buffer_size += n_tensors * 128 /* CLIP PADDING */;
 
     clip_ctx * new_clip = new clip_ctx;
+
 #ifdef GGML_USE_CUBLAS
     new_clip->backend = ggml_backend_cuda_init(0);
     printf("%s: CLIP using CUDA backend\n", __func__);
@@ -543,6 +527,7 @@ struct clip_ctx * clip_model_load(const char * fname, const int verbosity = 1) {
     printf("%s: CLIP using Metal backend\n", __func__);
 #endif
 
+
     if (!new_clip->backend) {
         new_clip->backend = ggml_backend_cpu_init();
         printf("%s: CLIP using CPU backend\n", __func__);
@@ -931,26 +916,8 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
 
     ggml_type type = GGML_TYPE_Q4_1;
 
-    switch (itype) {
-        case 2:
-            type = GGML_TYPE_Q4_0;
-            break;
-        case 3:
-            type = GGML_TYPE_Q4_1;
-            break;
-        case 6:
-            type = GGML_TYPE_Q5_0;
-            break;
-        case 7:
-            type = GGML_TYPE_Q5_1;
-            break;
-        case 8:
-            type = GGML_TYPE_Q8_0;
-            break;
-        default:
-            fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype);
-            return false;
-    };
+    assert(itype < GGML_TYPE_COUNT);
+    type = static_cast<ggml_type>(itype);
 
     auto * ctx_clip = clip_model_load(fname_inp, 2);
 
@@ -1010,6 +977,10 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
 
         if (quantize) {
             new_type = type;
+            if (new_type >= GGML_TYPE_Q2_K && name.find("embd") != std::string::npos) {
+                new_type = GGML_TYPE_Q8_0; // ggml_get_rows needs non K type
+                // fprintf(stderr, "%s: quantizing %s to %s\n", __func__, name.c_str(), ggml_type_name(new_type));
+            }
             const size_t n_elms = ggml_nelements(cur);
             float * f32_data;
 
@@ -1054,6 +1025,21 @@ bool clip_model_quantize(const char * fname_inp, const char * fname_out, const i
                 case GGML_TYPE_Q8_0: {
                     new_size = ggml_quantize_q8_0(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
                 } break;
+                case GGML_TYPE_Q2_K: {
+                    new_size = ggml_quantize_q2_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q3_K: {
+                    new_size = ggml_quantize_q3_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q4_K: {
+                    new_size = ggml_quantize_q4_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q5_K: {
+                    new_size = ggml_quantize_q5_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
+                case GGML_TYPE_Q6_K: {
+                    new_size = ggml_quantize_q6_K(f32_data, new_data, n_elms, cur->ne[0], hist_cur.data());
+                } break;
                 default: {
                     fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, new_type);
                     return false;

examples/llava/llava-cli.cpp

@@ -243,6 +243,9 @@ int main(int argc, char ** argv) {
     }
 
     auto image_embed = load_image(ctx_llava, &params);
+    if (!image_embed) {
+        return 1;
+    }
 
     // process the prompt
     process_prompt(ctx_llava, image_embed, &params, params.prompt);

examples/main/main.cpp

@@ -439,6 +439,21 @@ int main(int argc, char ** argv) {
     LOG_TEE("sampling: \n%s\n", llama_sampling_print(sparams).c_str());
     LOG_TEE("sampling order: \n%s\n", llama_sampling_order_print(sparams).c_str());
     LOG_TEE("generate: n_ctx = %d, n_batch = %d, n_predict = %d, n_keep = %d\n", n_ctx, params.n_batch, params.n_predict, params.n_keep);
+
+    // group-attention state
+    // number of grouped KV tokens so far (used only if params.grp_attn_n > 1)
+    int ga_i = 0;
+
+    const int ga_n = params.grp_attn_n;
+    const int ga_w = params.grp_attn_w;
+
+    if (ga_n != 1) {
+        GGML_ASSERT(ga_n > 0                    && "grp_attn_n must be positive");                     // NOLINT
+        GGML_ASSERT(ga_w % ga_n == 0            && "grp_attn_w must be a multiple of grp_attn_n");     // NOLINT
+      //GGML_ASSERT(n_ctx_train % ga_w == 0     && "n_ctx_train must be a multiple of grp_attn_w");    // NOLINT
+      //GGML_ASSERT(n_ctx >= n_ctx_train * ga_n && "n_ctx must be at least n_ctx_train * grp_attn_n"); // NOLINT
+        LOG_TEE("self-extend: n_ctx_train = %d, grp_attn_n = %d, grp_attn_w = %d\n", n_ctx_train, ga_n, ga_w);
+    }
     LOG_TEE("\n\n");
 
     if (params.interactive) {
@@ -485,7 +500,7 @@ int main(int argc, char ** argv) {
     while ((n_remain != 0 && !is_antiprompt) || params.interactive) {
         // predict
         if (!embd.empty()) {
-            // Note: n_ctx - 4 here is to match the logic for commandline prompt handling via
+            // Note: (n_ctx - 4) here is to match the logic for commandline prompt handling via
             // --prompt or --file which uses the same value.
             int max_embd_size = n_ctx - 4;
 
@@ -500,37 +515,61 @@ int main(int argc, char ** argv) {
                 fflush(stdout);
             }
 
-            // infinite text generation via context swapping
-            // if we run out of context:
-            // - take the n_keep first tokens from the original prompt (via n_past)
-            // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
-            if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
-                if (params.n_predict == -2) {
-                    LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
-                    break;
+            if (ga_n == 1) {
+                // infinite text generation via context shifting
+                // if we run out of context:
+                // - take the n_keep first tokens from the original prompt (via n_past)
+                // - take half of the last (n_ctx - n_keep) tokens and recompute the logits in batches
+                if (n_past + (int) embd.size() + std::max<int>(0, guidance_offset) > n_ctx) {
+                    if (params.n_predict == -2) {
+                        LOG_TEE("\n\n%s: context full and n_predict == -%d => stopping\n", __func__, params.n_predict);
+                        break;
+                    }
+
+                    const int n_left    = n_past - params.n_keep - 1;
+                    const int n_discard = n_left/2;
+
+                    LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
+                            n_past, n_left, n_ctx, params.n_keep, n_discard);
+
+                    llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
+                    llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+
+                    n_past -= n_discard;
+
+                    if (ctx_guidance) {
+                        n_past_guidance -= n_discard;
+                    }
+
+                    LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
+
+                    LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
+
+                    LOG("clear session path\n");
+                    path_session.clear();
                 }
+            } else {
+                // context extension via Self-Extend
+                while (n_past >= ga_i + ga_w) {
+                    const int ib = (ga_n*ga_i)/ga_w;
+                    const int bd = (ga_w/ga_n)*(ga_n - 1);
+                    const int dd = (ga_w/ga_n) - ib*bd - ga_w;
 
-                const int n_left    = n_past - params.n_keep - 1;
-                const int n_discard = n_left/2;
+                    LOG("\n");
+                    LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i, n_past, ib*bd, ga_i + ib*bd, n_past + ib*bd);
+                    LOG("div:   [%6d, %6d] / %6d -> [%6d, %6d]\n", ga_i + ib*bd, ga_i + ib*bd + ga_w, ga_n, (ga_i + ib*bd)/ga_n, (ga_i + ib*bd + ga_w)/ga_n);
+                    LOG("shift: [%6d, %6d] + %6d -> [%6d, %6d]\n", ga_i + ib*bd + ga_w, n_past + ib*bd, dd, ga_i + ib*bd + ga_w + dd, n_past + ib*bd + dd);
 
-                LOG("context full, swapping: n_past = %d, n_left = %d, n_ctx = %d, n_keep = %d, n_discard = %d\n",
-                    n_past, n_left, n_ctx, params.n_keep, n_discard);
+                    llama_kv_cache_seq_shift(ctx, 0, ga_i,                n_past,              ib*bd);
+                    llama_kv_cache_seq_div  (ctx, 0, ga_i + ib*bd,        ga_i + ib*bd + ga_w, ga_n);
+                    llama_kv_cache_seq_shift(ctx, 0, ga_i + ib*bd + ga_w, n_past + ib*bd,      dd);
 
-                llama_kv_cache_seq_rm   (ctx, 0, params.n_keep + 1            , params.n_keep + n_discard + 1);
-                llama_kv_cache_seq_shift(ctx, 0, params.n_keep + 1 + n_discard, n_past, -n_discard);
+                    n_past -= bd;
 
-                n_past -= n_discard;
+                    ga_i += ga_w/ga_n;
 
-                if (ctx_guidance) {
-                    n_past_guidance -= n_discard;
+                    LOG("\nn_past_old = %d, n_past = %d, ga_i = %d\n\n", n_past + bd, n_past, ga_i);
                 }
-
-                LOG("after swap: n_past = %d, n_past_guidance = %d\n", n_past, n_past_guidance);
-
-                LOG("embd: %s\n", LOG_TOKENS_TOSTR_PRETTY(ctx, embd).c_str());
-
-                LOG("clear session path\n");
-                path_session.clear();
             }
 
             // try to reuse a matching prefix from the loaded session instead of re-eval (via n_past)
@@ -611,6 +650,10 @@ int main(int argc, char ** argv) {
                 n_past += n_eval;
 
                 LOG("n_past = %d\n", n_past);
+                // Display total tokens alongside total time
+                if (params.n_print > 0 && n_past % params.n_print == 0) {
+                    LOG_TEE("\n\033[31mTokens consumed so far = %d / %d \033[0m\n", n_past, n_ctx);
+                }
             }
 
             if (!embd.empty() && !path_session.empty()) {

examples/passkey/CMakeLists.txt

@@ -0,0 +1,5 @@
+set(TARGET passkey)
+add_executable(${TARGET} passkey.cpp)
+install(TARGETS ${TARGET} RUNTIME)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

examples/passkey/README.md

@@ -0,0 +1,12 @@
+# llama.cpp/example/passkey
+
+See the following PRs for more info:
+
+- https://github.com/ggerganov/llama.cpp/pull/3856
+- https://github.com/ggerganov/llama.cpp/pull/4810
+
+### Usage
+
+```bash
+make -j && ./passkey ./models/llama-7b-v2/ggml-model-f16.gguf 250
+```

examples/passkey/passkey.cpp

@@ -0,0 +1,296 @@
+#include "common.h"
+#include "llama.h"
+
+#include <cmath>
+#include <cstdio>
+#include <string>
+#include <vector>
+
+int main(int argc, char ** argv) {
+    gpt_params params;
+
+    if (argc == 1 || argv[1][0] == '-') {
+        printf("usage: %s MODEL_PATH N_JUNK N_GRP I_POS SEED\n" , argv[0]);
+        return 1 ;
+    }
+
+    int seed = -1;
+
+    int n_junk = 250; // number of times to repeat the junk text
+    int n_keep = 32;  // number of tokens in the prompt prefix
+    int n_grp  = 1;   // if more than 1 - perform LongLM SelfExtend
+    int i_pos  = -1;  // position of the passkey in the junk text
+
+    if (argc >= 2) {
+        params.model = argv[1];
+    }
+
+    if (argc >= 3) {
+        n_junk = std::stoi(argv[2]);
+    }
+
+    if (argc >= 4) {
+        n_grp = std::stoi(argv[3]);
+    }
+
+    if (argc >= 5) {
+        i_pos = std::stoi(argv[4]);
+    }
+
+    if (argc >= 6) {
+        seed = std::stoi(argv[5]);
+    }
+
+    if (seed == -1) {
+        seed = time(NULL);
+    }
+
+    srand(seed);
+
+    if (i_pos == -1) {
+        i_pos = rand() % n_junk;
+    }
+
+    const std::string prompt_prefix = "There is an important info hidden inside a lot of irrelevant text. Find it and memorize them. I will quiz you about the important information there.";
+    const std::string prompt_suffix = " What is the pass key? The pass key is";
+
+    // generate junk text
+    params.prompt = prompt_prefix;
+
+    const int passkey = rand() % 50000 + 1;
+
+    for (int i = 0; i < n_junk; i++) {
+        if (i % n_junk == i_pos) {
+            params.prompt += " The pass key is " + std::to_string(passkey) + ". Remember it. " + std::to_string(passkey) + " is the pass key.";
+        }
+
+        params.prompt += " The grass is green. The sky is blue. The sun is yellow. Here we go. There and back again.";
+    }
+
+    params.prompt += prompt_suffix;
+
+    // init LLM
+
+    llama_backend_init(params.numa);
+
+    // initialize the model
+
+    llama_model_params model_params = llama_model_default_params();
+
+    model_params.n_gpu_layers = 99; // offload all layers to the GPU
+
+    llama_model * model = llama_load_model_from_file(params.model.c_str(), model_params);
+
+    if (model == NULL) {
+        fprintf(stderr , "%s: error: unable to load model\n" , __func__);
+        return 1;
+    }
+
+    // initialize the context
+
+    llama_context_params ctx_params = llama_context_default_params();
+
+    ctx_params.seed    = seed;
+    ctx_params.n_ctx   = llama_n_ctx_train(model)*n_grp + n_keep;
+    ctx_params.n_batch = 512;
+    ctx_params.n_threads       = params.n_threads;
+    ctx_params.n_threads_batch = params.n_threads_batch == -1 ? params.n_threads : params.n_threads_batch;
+
+    GGML_ASSERT(ctx_params.n_batch % n_grp == 0 && "n_batch must be divisible by n_grp");
+
+    llama_context * ctx = llama_new_context_with_model(model, ctx_params);
+
+    if (ctx == NULL) {
+        fprintf(stderr , "%s: error: failed to create the llama_context\n" , __func__);
+        return 1;
+    }
+
+    // tokenize the prompt
+    std::vector<llama_token> tokens_list;
+    tokens_list = ::llama_tokenize(ctx, params.prompt, true);
+
+    // tokenize the prefix and use it as a sink
+    const int n_tokens_prefix = ::llama_tokenize(ctx, prompt_prefix, true).size();
+
+    const int n_tokens_all = tokens_list.size();
+
+    // we leave a margin of 16 tokens for the generated text - it should contain just the passkey
+    const int n_predict = 16;
+
+    // total length of the sequences including the prompt
+    const int n_len = n_tokens_all + n_predict;
+
+    const int n_ctx       = llama_n_ctx(ctx) - n_keep;
+    const int n_kv_req    = llama_n_ctx(ctx);
+    const int n_batch     = ctx_params.n_batch;
+    const int n_batch_grp = ctx_params.n_batch/n_grp;
+
+    LOG_TEE("\n%s: n_len = %d, n_ctx = %d, n_kv_req = %d, n_grp = %d, n_batch = %d\n", __func__, n_len, n_ctx, n_kv_req, n_grp, n_batch);
+
+    // print the prompt token-by-token
+
+    LOG_TEE("\n");
+    LOG_TEE("prefix tokens: %d\n", n_tokens_prefix);
+    LOG_TEE("prompt tokens: %d\n", n_tokens_all);
+    //LOG_TEE("prompt: %s\n", params.prompt.c_str());
+
+    llama_batch batch = llama_batch_init(512, 0, 1);
+
+    int n_past = 0;
+
+    // fill the KV cache
+    for (int i = 0; i < n_ctx; i += n_batch) {
+        if (i > 0 && n_grp > 1) {
+            // if SelfExtend is enabled, we compress the position from the last batch by a factor of n_grp
+            const int ib = i/n_batch - 1;
+            const int bd = n_batch_grp*(n_grp - 1);
+
+            llama_kv_cache_seq_shift(ctx, 0, n_past - n_batch,         n_past,         ib*bd);
+            llama_kv_cache_seq_div  (ctx, 0, n_past - n_batch + ib*bd, n_past + ib*bd, n_grp);
+
+            n_past -= bd;
+        }
+
+        llama_batch_clear(batch);
+
+        for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
+            llama_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
+        }
+
+        if (i + n_batch >= n_tokens_all) {
+            batch.logits[batch.n_tokens - 1] = true;
+        }
+
+        if (llama_decode(ctx, batch) != 0) {
+            LOG_TEE("%s: llama_decode() failed\n", __func__);
+            return 1;
+        }
+
+        LOG_TEE("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
+
+        if (i + n_batch >= n_tokens_all) {
+            break;
+        }
+    }
+
+    for (int i = n_ctx; i < n_tokens_all; i += n_batch) {
+        const int n_discard = n_batch;
+
+        LOG_TEE("%s: shifting KV cache with %d\n", __func__, n_discard);
+
+        llama_kv_cache_seq_rm   (ctx, 0, n_keep            , n_keep + n_discard);
+        llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+
+        n_past -= n_discard;
+
+        llama_batch_clear(batch);
+
+        for (int j = 0; j < n_batch && i + j < n_tokens_all; j++) {
+            llama_batch_add(batch, tokens_list[i + j], n_past++, { 0 }, false);
+        }
+
+        if (i + n_batch >= n_tokens_all) {
+            batch.logits[batch.n_tokens - 1] = true;
+        }
+
+        if (llama_decode(ctx, batch) != 0) {
+            LOG_TEE("%s: llama_decode() failed\n", __func__);
+            return 1;
+        }
+
+        LOG_TEE("%s: processed: [%6d, %6d)\n", __func__, i, std::min(i + n_batch, n_tokens_all));
+    }
+
+    {
+        const int n_discard = n_past - n_ctx + n_predict;
+
+        if (n_discard > 0) {
+            LOG_TEE("%s: shifting KV cache with %d to free space for the answer\n", __func__, n_discard);
+
+            llama_kv_cache_seq_rm   (ctx, 0, n_keep            , n_keep + n_discard);
+            llama_kv_cache_seq_shift(ctx, 0, n_keep + n_discard, n_ctx,  -n_discard);
+
+            n_past -= n_discard;
+        }
+    }
+
+    LOG_TEE("\n");
+    LOG_TEE("%s: passkey = %d, inserted at position %d / %d (token pos: ~%d)\n", __func__, passkey, i_pos, n_junk, (i_pos * n_tokens_all) / n_junk);
+    LOG_TEE("\n");
+
+    // main loop
+
+    int n_cur    = n_tokens_all;
+    int n_decode = 0;
+
+    LOG_TEE("%s", prompt_suffix.c_str());
+    fflush(stdout);
+
+    const auto t_main_start = ggml_time_us();
+
+    while (n_cur <= n_len) {
+        // sample the next token
+        {
+            auto   n_vocab = llama_n_vocab(model);
+            auto * logits  = llama_get_logits_ith(ctx, batch.n_tokens - 1);
+
+            std::vector<llama_token_data> candidates;
+            candidates.reserve(n_vocab);
+
+            for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
+                candidates.emplace_back(llama_token_data{ token_id, logits[token_id], 0.0f });
+            }
+
+            llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
+
+            // sample the most likely token
+            const llama_token new_token_id = llama_sample_token_greedy(ctx, &candidates_p);
+
+            // is it an end of stream?
+            if (new_token_id == llama_token_eos(model) || n_cur == n_len) {
+                LOG_TEE("\n");
+
+                break;
+            }
+
+            LOG_TEE("%s", llama_token_to_piece(ctx, new_token_id).c_str());
+            fflush(stdout);
+
+            n_decode += 1;
+
+            // prepare the next batch
+            llama_batch_clear(batch);
+
+            // push this new token for next evaluation
+            llama_batch_add(batch, new_token_id, n_past++, { 0 }, true);
+        }
+
+        n_cur += 1;
+
+        // evaluate the current batch with the transformer model
+        if (llama_decode(ctx, batch)) {
+            fprintf(stderr, "%s : failed to eval, return code %d\n", __func__, 1);
+            return 1;
+        }
+    }
+
+    LOG_TEE("\n");
+
+    const auto t_main_end = ggml_time_us();
+
+    LOG_TEE("%s: decoded %d tokens in %.2f s, speed: %.2f t/s\n",
+            __func__, n_decode, (t_main_end - t_main_start) / 1000000.0f, n_decode / ((t_main_end - t_main_start) / 1000000.0f));
+
+    llama_print_timings(ctx);
+
+    fprintf(stderr, "\n");
+
+    llama_batch_free(batch);
+
+    llama_free(ctx);
+    llama_free_model(model);
+
+    llama_backend_free();
+
+    return 0;
+}

examples/quantize/quantize.cpp

@@ -18,6 +18,7 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
     { "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", },
     { "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" },
     { "Q3_K_S", LLAMA_FTYPE_MOSTLY_Q3_K_S, " 2.75G, +0.5551 ppl @ LLaMA-v1-7B", },
     { "Q3_K_M", LLAMA_FTYPE_MOSTLY_Q3_K_M, " 3.07G, +0.2496 ppl @ LLaMA-v1-7B", },

examples/server/README.md

@@ -23,6 +23,8 @@ Command line options:
 -   `--host`: Set the hostname or ip address to listen. Default `127.0.0.1`.
 -   `--port`: Set the port to listen. Default: `8080`.
 -   `--path`: path from which to serve static files (default examples/server/public)
+-   `--api-key`: Set an api key for request authorization. By default the server responds to every request. With an api key set, the requests must have the Authorization header set with the api key as Bearer token. May be used multiple times to enable multiple valid keys.
+-   `--api-key-file`: path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access. May be used in conjunction with `--api-key`'s.
 -   `--embedding`: Enable embedding extraction, Default: disabled.
 -   `-np N`, `--parallel N`: Set the number of slots for process requests (default: 1)
 -   `-cb`, `--cont-batching`: enable continuous batching (a.k.a dynamic batching) (default: disabled)
@@ -109,6 +111,10 @@ node index.js
 ```
 
 ## API Endpoints
+- **GET** `/health`: Returns the current state of the server:
+    - `{"status": "loading model"}` if the model is still being loaded.
+    - `{"status": "error"}` if the model failed to load.
+    - `{"status": "ok"}` if the model is successfully loaded and the server is ready for further requests mentioned below.
 
 -   **POST** `/completion`: Given a `prompt`, it returns the predicted completion.
 
@@ -174,35 +180,44 @@ node index.js
 
     `system_prompt`: Change the system prompt (initial prompt of all slots), this is useful for chat applications. [See more](#change-system-prompt-on-runtime)
 
-    *Result JSON:*
+### Result JSON:
 
-    Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
+* Note: When using streaming mode (`stream`) only `content` and `stop` will be returned until end of completion.
 
-    `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
 
-    `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
+- `completion_probabilities`: An array of token probabilities for each completion. The array's length is `n_predict`. Each item in the array has the following structure:
 
-    `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`
+```
+{
+  "content": "<the token selected by the model>",
+  "probs": [
+    {
+      "prob": float,
+      "tok_str": "<most likely token>"
+    },
+    {
+      "prob": float,
+      "tok_str": "<second most likely tonen>"
+    },
+    ...
+  ]
+},
+```
+Notice that each `probs` is an array of length `n_probs`.
 
-    `model`: The path to the model loaded with `-m`
-
-    `prompt`: The provided `prompt`
-
-    `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token
-
-    `stopped_limit`: Indicating whether the completion stopped because `n_predict` tokens were generated before stop words or EOS was encountered
-
-    `stopped_word`: Indicating whether the completion stopped due to encountering a stopping word from `stop` JSON array provided
-
-    `stopping_word`: The stopping word encountered which stopped the generation (or "" if not stopped due to a stopping word)
-
-    `timings`: Hash of timing information about the completion such as the number of tokens `predicted_per_second`
-
-    `tokens_cached`: Number of tokens from the prompt which could be re-used from previous completion (`n_past`)
-
-    `tokens_evaluated`: Number of tokens evaluated in total from the prompt
-
-    `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
+- `content`: Completion result as a string (excluding `stopping_word` if any). In case of streaming mode, will contain the next token as a string.
+- `stop`: Boolean for use with `stream` to check whether the generation has stopped (Note: This is not related to stopping words array `stop` from input options)
+- `generation_settings`: The provided options above excluding `prompt` but including `n_ctx`, `model`
+- `model`: The path to the model loaded with `-m`
+- `prompt`: The provided `prompt`
+- `stopped_eos`: Indicating whether the completion has stopped because it encountered the EOS token
+- `stopped_limit`: Indicating whether the completion stopped because `n_predict` tokens were generated before stop words or EOS was encountered
+- `stopped_word`: Indicating whether the completion stopped due to encountering a stopping word from `stop` JSON array provided
+- `stopping_word`: The stopping word encountered which stopped the generation (or "" if not stopped due to a stopping word)
+- `timings`: Hash of timing information about the completion such as the number of tokens `predicted_per_second`
+- `tokens_cached`: Number of tokens from the prompt which could be re-used from previous completion (`n_past`)
+- `tokens_evaluated`: Number of tokens evaluated in total from the prompt
+- `truncated`: Boolean indicating if the context size was exceeded during generation, i.e. the number of tokens provided in the prompt (`tokens_evaluated`) plus tokens generated (`tokens predicted`) exceeded the context size (`n_ctx`)
 
 -   **POST** `/tokenize`: Tokenize a given text.
 

examples/server/server.cpp

@@ -26,6 +26,7 @@
 #include <mutex>
 #include <chrono>
 #include <condition_variable>
+#include <atomic>
 
 #ifndef SERVER_VERBOSE
 #define SERVER_VERBOSE 1
@@ -38,7 +39,7 @@ using json = nlohmann::json;
 struct server_params
 {
     std::string hostname = "127.0.0.1";
-    std::string api_key;
+    std::vector<std::string> api_keys;
     std::string public_path = "examples/server/public";
     int32_t port = 8080;
     int32_t read_timeout = 600;
@@ -146,9 +147,15 @@ static std::vector<uint8_t> base64_decode(const std::string & encoded_string)
 // parallel
 //
 
+enum server_state {
+    SERVER_STATE_LOADING_MODEL,  // Server is starting up, model not fully loaded yet
+    SERVER_STATE_READY,          // Server is ready and model is loaded
+    SERVER_STATE_ERROR           // An error occurred, load_model failed
+};
+
 enum task_type {
-    COMPLETION_TASK,
-    CANCEL_TASK
+    TASK_TYPE_COMPLETION,
+    TASK_TYPE_CANCEL,
 };
 
 struct task_server {
@@ -1395,11 +1402,11 @@ struct llama_server_context
         task.data = std::move(data);
         task.infill_mode = infill;
         task.embedding_mode = embedding;
-        task.type = COMPLETION_TASK;
+        task.type = TASK_TYPE_COMPLETION;
         task.multitask_id = multitask_id;
 
         // when a completion task's prompt array is not a singleton, we split it into multiple requests
-        if (task.data.at("prompt").size() > 1)
+        if (task.data.count("prompt") && task.data.at("prompt").size() > 1)
         {
             lock.unlock(); // entering new func scope
             return split_multiprompt_task(task);
@@ -1517,7 +1524,7 @@ struct llama_server_context
         std::unique_lock<std::mutex> lock(mutex_tasks);
         task_server task;
         task.id = id_gen++;
-        task.type = CANCEL_TASK;
+        task.type = TASK_TYPE_CANCEL;
         task.target_id = task_id;
         queue_tasks.push_back(task);
         condition_tasks.notify_one();
@@ -1553,7 +1560,7 @@ struct llama_server_context
             queue_tasks.erase(queue_tasks.begin());
             switch (task.type)
             {
-                case COMPLETION_TASK: {
+                case TASK_TYPE_COMPLETION: {
                     llama_client_slot *slot = get_slot(json_value(task.data, "slot_id", -1));
                     if (slot == nullptr)
                     {
@@ -1570,9 +1577,9 @@ struct llama_server_context
 
                     slot->reset();
 
-                    slot->infill = task.infill_mode;
-                    slot->embedding = task.embedding_mode;
-                    slot->task_id = task.id;
+                    slot->infill       = task.infill_mode;
+                    slot->embedding    = task.embedding_mode;
+                    slot->task_id      = task.id;
                     slot->multitask_id = task.multitask_id;
 
                     if (!launch_slot_with_data(slot, task.data))
@@ -1582,7 +1589,7 @@ struct llama_server_context
                         break;
                     }
                 } break;
-                case CANCEL_TASK: { // release slot linked with the task id
+                case TASK_TYPE_CANCEL: { // release slot linked with the task id
                     for (auto & slot : slots)
                     {
                         if (slot.task_id == task.target_id)
@@ -1724,7 +1731,8 @@ struct llama_server_context
                 const bool has_prompt = slot.prompt.is_array() || (slot.prompt.is_string() && !slot.prompt.get<std::string>().empty()) || !slot.images.empty();
 
                 // empty prompt passed -> release the slot and send empty response
-                if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt)
+                // note: infill mode allows empty prompt
+                if (slot.state == IDLE && slot.command == LOAD_PROMPT && !has_prompt && !slot.infill)
                 {
                     slot.release();
                     slot.print_timings();
@@ -2014,6 +2022,7 @@ static void server_print_usage(const char *argv0, const gpt_params &params,
     printf("  --port PORT           port to listen (default  (default: %d)\n", sparams.port);
     printf("  --path PUBLIC_PATH    path from which to serve static files (default %s)\n", sparams.public_path.c_str());
     printf("  --api-key API_KEY     optional api key to enhance server security. If set, requests must include this key for access.\n");
+    printf("  --api-key-file FNAME  path to file containing api keys delimited by new lines. If set, requests must include one of the keys for access.\n");
     printf("  -to N, --timeout N    server read/write timeout in seconds (default: %d)\n", sparams.read_timeout);
     printf("  --embedding           enable embedding vector output (default: %s)\n", params.embedding ? "enabled" : "disabled");
     printf("  -np N, --parallel N   number of slots for process requests (default: %d)\n", params.n_parallel);
@@ -2074,7 +2083,28 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
                 invalid_param = true;
                 break;
             }
-            sparams.api_key = argv[i];
+            sparams.api_keys.push_back(argv[i]);
+        }
+        else if (arg == "--api-key-file")
+        {
+            if (++i >= argc)
+            {
+                invalid_param = true;
+                break;
+            }
+            std::ifstream key_file(argv[i]);
+            if (!key_file) {
+                fprintf(stderr, "error: failed to open file '%s'\n", argv[i]);
+                invalid_param = true;
+                break;
+            }
+            std::string key;
+            while (std::getline(key_file, key)) {
+               if (key.size() > 0) {
+                   sparams.api_keys.push_back(key);
+               }
+            }
+            key_file.close();
         }
         else if (arg == "--timeout" || arg == "-to")
         {
@@ -2453,7 +2483,6 @@ static void server_params_parse(int argc, char **argv, server_params &sparams,
     }
 }
 
-
 static std::string random_string()
 {
     static const std::string str("0123456789ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghijklmnopqrstuvwxyz");
@@ -2509,7 +2538,7 @@ json oaicompat_completion_params_parse(
     //
     // https://platform.openai.com/docs/api-reference/chat/create
     llama_sampling_params default_sparams;
-    llama_params["model"]             = json_value(body, "model", std::string("uknown"));
+    llama_params["model"]             = json_value(body, "model", std::string("unknown"));
     llama_params["prompt"]            = format_chatml(body["messages"]); // OpenAI 'messages' to llama.cpp 'prompt'
     llama_params["cache_prompt"]      = json_value(body, "cache_prompt", false);
     llama_params["temperature"]       = json_value(body, "temperature", 0.0);
@@ -2581,8 +2610,8 @@ static json format_final_response_oaicompat(const json &request, const task_resu
             {"object", streaming ? "chat.completion.chunk" : "chat.completion"},
             {"usage",
                 json{{"completion_tokens", num_tokens_predicted},
-                    {"prompt_tokens", num_prompt_tokens},
-                    {"total_tokens", num_tokens_predicted + num_prompt_tokens}}},
+                     {"prompt_tokens",     num_prompt_tokens},
+                     {"total_tokens",      num_tokens_predicted + num_prompt_tokens}}},
             {"id", gen_chatcmplid()}};
 
     if (server_verbose) {
@@ -2790,20 +2819,131 @@ int main(int argc, char **argv)
                                 {"system_info", llama_print_system_info()},
                             });
 
-    // load the model
-    if (!llama.load_model(params))
+    httplib::Server svr;
+
+    std::atomic<server_state> state{SERVER_STATE_LOADING_MODEL};
+
+    svr.set_default_headers({{"Server", "llama.cpp"}});
+
+    // CORS preflight
+    svr.Options(R"(.*)", [](const httplib::Request &req, httplib::Response &res) {
+        res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
+        res.set_header("Access-Control-Allow-Credentials", "true");
+        res.set_header("Access-Control-Allow-Methods", "POST");
+        res.set_header("Access-Control-Allow-Headers", "*");
+    });
+
+    svr.Get("/health", [&](const httplib::Request&, httplib::Response& res) {
+        server_state current_state = state.load();
+        switch(current_state) {
+            case SERVER_STATE_READY:
+                res.set_content(R"({"status": "ok"})", "application/json");
+                res.status = 200; // HTTP OK
+                break;
+            case SERVER_STATE_LOADING_MODEL:
+                res.set_content(R"({"status": "loading model"})", "application/json");
+                res.status = 503; // HTTP Service Unavailable
+                break;
+            case SERVER_STATE_ERROR:
+                res.set_content(R"({"status": "error", "error": "Model failed to load"})", "application/json");
+                res.status = 500; // HTTP Internal Server Error
+                break;
+        }
+    });
+
+    svr.set_logger(log_server_request);
+
+    svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep)
+            {
+                const char fmt[] = "500 Internal Server Error\n%s";
+                char buf[BUFSIZ];
+                try
+                {
+                    std::rethrow_exception(std::move(ep));
+                }
+                catch (std::exception &e)
+                {
+                    snprintf(buf, sizeof(buf), fmt, e.what());
+                }
+                catch (...)
+                {
+                    snprintf(buf, sizeof(buf), fmt, "Unknown Exception");
+                }
+                res.set_content(buf, "text/plain; charset=utf-8");
+                res.status = 500;
+            });
+
+    svr.set_error_handler([](const httplib::Request &, httplib::Response &res)
+            {
+                if (res.status == 401)
+                {
+                    res.set_content("Unauthorized", "text/plain; charset=utf-8");
+                }
+                if (res.status == 400)
+                {
+                    res.set_content("Invalid request", "text/plain; charset=utf-8");
+                }
+                else if (res.status == 404)
+                {
+                    res.set_content("File Not Found", "text/plain; charset=utf-8");
+                    res.status = 404;
+                }
+            });
+
+    // set timeouts and change hostname and port
+    svr.set_read_timeout (sparams.read_timeout);
+    svr.set_write_timeout(sparams.write_timeout);
+
+    if (!svr.bind_to_port(sparams.hostname, sparams.port))
     {
+        fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port);
         return 1;
     }
 
-    llama.initialize();
+    // Set the base directory for serving static files
+    svr.set_base_dir(sparams.public_path);
 
-    httplib::Server svr;
+    // to make it ctrl+clickable:
+    LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
+
+    std::unordered_map<std::string, std::string> log_data;
+    log_data["hostname"] = sparams.hostname;
+    log_data["port"] = std::to_string(sparams.port);
+
+    if (sparams.api_keys.size() == 1) {
+        log_data["api_key"] = "api_key: ****" + sparams.api_keys[0].substr(sparams.api_keys[0].length() - 4);
+    } else if (sparams.api_keys.size() > 1) {
+        log_data["api_key"] = "api_key: " + std::to_string(sparams.api_keys.size()) + " keys loaded";
+    }
+
+    LOG_INFO("HTTP server listening", log_data);
+    // run the HTTP server in a thread - see comment below
+    std::thread t([&]()
+            {
+                if (!svr.listen_after_bind())
+                {
+                    state.store(SERVER_STATE_ERROR);
+                    return 1;
+                }
+
+                return 0;
+            });
+
+    // load the model
+    if (!llama.load_model(params))
+    {
+        state.store(SERVER_STATE_ERROR);
+        return 1;
+    } else {
+        llama.initialize();
+        state.store(SERVER_STATE_READY);
+        LOG_INFO("model loaded", {});
+    }
 
     // Middleware for API key validation
     auto validate_api_key = [&sparams](const httplib::Request &req, httplib::Response &res) -> bool {
         // If API key is not set, skip validation
-        if (sparams.api_key.empty()) {
+        if (sparams.api_keys.empty()) {
             return true;
         }
 
@@ -2812,7 +2952,7 @@ int main(int argc, char **argv)
         std::string prefix = "Bearer ";
         if (auth_header.substr(0, prefix.size()) == prefix) {
             std::string received_api_key = auth_header.substr(prefix.size());
-            if (received_api_key == sparams.api_key) {
+            if (std::find(sparams.api_keys.begin(), sparams.api_keys.end(), received_api_key) != sparams.api_keys.end()) {
                 return true; // API key is valid
             }
         }
@@ -2826,10 +2966,6 @@ int main(int argc, char **argv)
         return false;
     };
 
-    svr.set_default_headers({{"Server", "llama.cpp"},
-                             {"Access-Control-Allow-Origin", "*"},
-                             {"Access-Control-Allow-Headers", "content-type"}});
-
     // this is only called if no index.html is found in the public --path
     svr.Get("/", [](const httplib::Request &, httplib::Response &res)
             {
@@ -2858,9 +2994,9 @@ int main(int argc, char **argv)
                 return false;
             });
 
-    svr.Get("/props", [&llama](const httplib::Request & /*req*/, httplib::Response &res)
+    svr.Get("/props", [&llama](const httplib::Request & req, httplib::Response &res)
             {
-                res.set_header("Access-Control-Allow-Origin", "*");
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 json data = {
                     { "user_name",      llama.name_user.c_str() },
                     { "assistant_name", llama.name_assistant.c_str() }
@@ -2870,6 +3006,7 @@ int main(int argc, char **argv)
 
     svr.Post("/completion", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 if (!validate_api_key(req, res)) {
                     return;
                 }
@@ -2937,10 +3074,9 @@ int main(int argc, char **argv)
                 }
             });
 
-
-
-    svr.Get("/v1/models", [&params](const httplib::Request&, httplib::Response& res)
+    svr.Get("/v1/models", [&params](const httplib::Request& req, httplib::Response& res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 std::time_t t = std::time(0);
 
                 json models = {
@@ -2958,9 +3094,11 @@ int main(int argc, char **argv)
                 res.set_content(models.dump(), "application/json; charset=utf-8");
             });
 
+
     // TODO: add mount point without "/v1" prefix -- how?
     svr.Post("/v1/chat/completions", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 if (!validate_api_key(req, res)) {
                     return;
                 }
@@ -3034,6 +3172,7 @@ int main(int argc, char **argv)
 
     svr.Post("/infill", [&llama, &validate_api_key](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 if (!validate_api_key(req, res)) {
                     return;
                 }
@@ -3106,6 +3245,7 @@ int main(int argc, char **argv)
 
     svr.Post("/tokenize", [&llama](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 const json body = json::parse(req.body);
                 std::vector<llama_token> tokens;
                 if (body.count("content") != 0)
@@ -3118,6 +3258,7 @@ int main(int argc, char **argv)
 
     svr.Post("/detokenize", [&llama](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 const json body = json::parse(req.body);
                 std::string content;
                 if (body.count("tokens") != 0)
@@ -3132,6 +3273,7 @@ int main(int argc, char **argv)
 
     svr.Post("/embedding", [&llama](const httplib::Request &req, httplib::Response &res)
             {
+                res.set_header("Access-Control-Allow-Origin", req.get_header_value("Origin"));
                 const json body = json::parse(req.body);
                 json prompt;
                 if (body.count("content") != 0)
@@ -3157,81 +3299,6 @@ int main(int argc, char **argv)
                 return res.set_content(result.result_json.dump(), "application/json; charset=utf-8");
             });
 
-    svr.set_logger(log_server_request);
-
-    svr.set_exception_handler([](const httplib::Request &, httplib::Response &res, std::exception_ptr ep)
-            {
-                const char fmt[] = "500 Internal Server Error\n%s";
-                char buf[BUFSIZ];
-                try
-                {
-                    std::rethrow_exception(std::move(ep));
-                }
-                catch (std::exception &e)
-                {
-                    snprintf(buf, sizeof(buf), fmt, e.what());
-                }
-                catch (...)
-                {
-                    snprintf(buf, sizeof(buf), fmt, "Unknown Exception");
-                }
-                res.set_content(buf, "text/plain; charset=utf-8");
-                res.status = 500;
-            });
-
-    svr.set_error_handler([](const httplib::Request &, httplib::Response &res)
-            {
-                if (res.status == 401)
-                {
-                    res.set_content("Unauthorized", "text/plain; charset=utf-8");
-                }
-                if (res.status == 400)
-                {
-                    res.set_content("Invalid request", "text/plain; charset=utf-8");
-                }
-                else if (res.status == 404)
-                {
-                    res.set_content("File Not Found", "text/plain; charset=utf-8");
-                    res.status = 404;
-                }
-            });
-
-    // set timeouts and change hostname and port
-    svr.set_read_timeout (sparams.read_timeout);
-    svr.set_write_timeout(sparams.write_timeout);
-
-    if (!svr.bind_to_port(sparams.hostname, sparams.port))
-    {
-        fprintf(stderr, "\ncouldn't bind to server socket: hostname=%s port=%d\n\n", sparams.hostname.c_str(), sparams.port);
-        return 1;
-    }
-
-    // Set the base directory for serving static files
-    svr.set_base_dir(sparams.public_path);
-
-    // to make it ctrl+clickable:
-    LOG_TEE("\nllama server listening at http://%s:%d\n\n", sparams.hostname.c_str(), sparams.port);
-
-    std::unordered_map<std::string, std::string> log_data;
-    log_data["hostname"] = sparams.hostname;
-    log_data["port"] = std::to_string(sparams.port);
-
-    if (!sparams.api_key.empty()) {
-        log_data["api_key"] = "api_key: ****" + sparams.api_key.substr(sparams.api_key.length() - 4);
-    }
-
-    LOG_INFO("HTTP server listening", log_data);
-    // run the HTTP server in a thread - see comment below
-    std::thread t([&]()
-            {
-                if (!svr.listen_after_bind())
-                {
-                    return 1;
-                }
-
-                return 0;
-            });
-
     // GG: if I put the main loop inside a thread, it crashes on the first request when build in Debug!?
     //     "Bus error: 10" - this is on macOS, it does not crash on Linux
     //std::thread t2([&]()

ggml-cuda.cu

@@ -116,6 +116,7 @@
 #include "ggml.h"
 #include "ggml-backend-impl.h"
 
+#define CC_PASCAL     600
 #define MIN_CC_DP4A   610 // minimum compute capability for __dp4a, an intrinsic for byte-wise dot products
 #define CC_VOLTA      700
 #define CC_OFFSET_AMD 1000000
@@ -477,6 +478,23 @@ typedef struct {
 } block_q6_K;
 static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_K block size/padding");
 
+#define QR2_XXS 8
+#define QI2_XXS (QK_K / (4*QR2_XXS))
+typedef struct {
+    half d;
+    uint16_t qs[QK_K/8];
+} block_iq2_xxs;
+static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
+
+#define QR2_XS 8
+#define QI2_XS (QK_K / (4*QR2_XS))
+typedef struct {
+    half d;
+    uint16_t qs[QK_K/8];
+    uint8_t  scales[QK_K/32];
+} block_iq2_xs;
+static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
+
 #define WARP_SIZE 32
 #define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
 
@@ -548,11 +566,12 @@ static float g_tensor_split[GGML_CUDA_MAX_DEVICES] = {0};
 
 struct cuda_device_capabilities {
     int     cc;                 // compute capability
+    size_t  smpb;               // max. shared memory per block
     bool    vmm;                // virtual memory support
     size_t  vmm_granularity;    // granularity of virtual memory
 };
 
-static cuda_device_capabilities g_device_caps[GGML_CUDA_MAX_DEVICES] = { {0, false, 0} };
+static cuda_device_capabilities g_device_caps[GGML_CUDA_MAX_DEVICES] = { {0, 0, false, 0} };
 
 static void * g_scratch_buffer = nullptr;
 static size_t g_scratch_size = 0; // disabled by default
@@ -585,6 +604,19 @@ static __device__ __forceinline__ float2 warp_reduce_sum(float2 a) {
     return a;
 }
 
+static __device__ __forceinline__ half2 warp_reduce_sum(half2 a) {
+#if __CUDA_ARCH__ < CC_PASCAL || (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+    (void) a;
+    bad_arch();
+#else
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        a = __hadd2(a, __shfl_xor_sync(0xffffffff, a, mask, 32));
+    }
+    return a;
+#endif // __CUDA_ARCH__ < CC_PASCAL || (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+}
+
 static __device__ __forceinline__ float warp_reduce_max(float x) {
 #pragma unroll
     for (int mask = 16; mask > 0; mask >>= 1) {
@@ -593,6 +625,19 @@ static __device__ __forceinline__ float warp_reduce_max(float x) {
     return x;
 }
 
+static __device__ __forceinline__ half2 warp_reduce_max(half2 x) {
+#if __CUDA_ARCH__ < CC_PASCAL || (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+    (void) x;
+    bad_arch();
+#else
+#pragma unroll
+    for (int mask = 16; mask > 0; mask >>= 1) {
+        x = __hmax2(x, __shfl_xor_sync(0xffffffff, x, mask, 32));
+    }
+    return x;
+#endif // __CUDA_ARCH__ < CC_PASCAL || (defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__))
+}
+
 static __device__ __forceinline__ float op_repeat(const float a, const float b) {
     return b;
     GGML_UNUSED(a);
@@ -1292,6 +1337,281 @@ static __global__ void dequantize_block_q6_K(const void * __restrict__ vx, dst_t
 #endif
 }
 
+static const __device__ uint64_t iq2xxs_grid[256] = {
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808,
+    0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819,
+    0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819,
+    0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b,
+    0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808,
+    0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08,
+    0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b,
+    0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819,
+    0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08,
+    0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808,
+    0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08,
+    0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808,
+    0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808,
+    0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919,
+    0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819,
+    0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08,
+    0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908,
+    0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819,
+    0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808,
+    0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808,
+    0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908,
+    0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808,
+    0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08,
+    0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819,
+    0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819,
+    0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819,
+    0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908,
+    0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19,
+    0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819,
+    0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b,
+    0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808,
+    0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908,
+    0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08,
+    0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08,
+    0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908,
+    0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819,
+    0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808,
+    0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808,
+    0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19,
+    0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819,
+    0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919,
+    0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b,
+    0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08,
+    0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808,
+    0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908,
+    0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b,
+    0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819,
+    0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08,
+    0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08,
+    0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808,
+    0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b,
+    0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b,
+    0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908,
+    0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819,
+    0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808,
+    0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908,
+    0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b,
+    0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808,
+    0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b,
+    0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b,
+    0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808,
+    0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19,
+    0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908,
+};
+
+static const __device__ uint64_t iq2xs_grid[512] = {
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
+    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
+    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
+    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
+    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808,
+    0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819,
+    0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819,
+    0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808,
+    0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b,
+    0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b,
+    0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908,
+    0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908,
+    0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919,
+    0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808,
+    0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919,
+    0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908,
+    0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b,
+    0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908,
+    0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08,
+    0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808,
+    0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808,
+    0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819,
+    0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908,
+    0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819,
+    0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808,
+    0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b,
+    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819,
+    0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819,
+    0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808,
+    0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908,
+    0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19,
+    0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b,
+    0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b,
+    0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919,
+    0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808,
+    0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819,
+    0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819,
+    0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b,
+    0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908,
+    0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808,
+    0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819,
+    0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808,
+    0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919,
+    0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808,
+    0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808,
+    0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908,
+    0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908,
+    0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808,
+    0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b,
+    0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819,
+    0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919,
+    0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908,
+    0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808,
+    0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908,
+    0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919,
+    0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08,
+    0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19,
+    0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b,
+    0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b,
+    0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808,
+    0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08,
+    0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b,
+    0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908,
+    0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b,
+    0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908,
+    0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08,
+    0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808,
+    0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808,
+    0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08,
+    0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819,
+    0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919,
+    0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808,
+    0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808,
+    0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819,
+    0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819,
+    0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908,
+    0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908,
+    0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b,
+    0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908,
+    0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908,
+    0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908,
+    0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808,
+    0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819,
+    0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819,
+    0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819,
+    0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808,
+    0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b,
+    0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819,
+    0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819,
+    0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08,
+    0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808,
+    0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19,
+    0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919,
+    0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808,
+    0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19,
+    0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b,
+    0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808,
+    0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b,
+    0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b,
+    0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08,
+    0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b,
+    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808,
+    0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819,
+    0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808,
+    0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808,
+    0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08,
+    0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b,
+    0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19,
+    0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08,
+    0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919,
+    0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08,
+    0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08,
+    0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908,
+    0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908,
+    0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b,
+    0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908,
+    0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808,
+    0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b,
+    0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808,
+    0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808,
+    0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19,
+    0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08,
+    0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808,
+    0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b,
+    0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808,
+    0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b,
+    0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
+};
+
+static const __device__ uint8_t ksigns_iq2xs[128] = {
+      0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
+    144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
+    160,  33,  34, 163,  36, 165, 166,  39,  40, 169, 170,  43, 172,  45,  46, 175,
+     48, 177, 178,  51, 180,  53,  54, 183, 184,  57,  58, 187,  60, 189, 190,  63,
+    192,  65,  66, 195,  68, 197, 198,  71,  72, 201, 202,  75, 204,  77,  78, 207,
+     80, 209, 210,  83, 212,  85,  86, 215, 216,  89,  90, 219,  92, 221, 222,  95,
+     96, 225, 226,  99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111,
+    240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
+};
+
+static const __device__ uint8_t kmask_iq2xs[8] = {1, 2, 4, 8, 16, 32, 64, 128};
+
+inline bool ggml_cuda_supports_mmq(enum ggml_type type) {
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+        case GGML_TYPE_Q4_1:
+        case GGML_TYPE_Q5_0:
+        case GGML_TYPE_Q5_1:
+        case GGML_TYPE_Q8_0:
+        case GGML_TYPE_Q2_K:
+        case GGML_TYPE_Q3_K:
+        case GGML_TYPE_Q4_K:
+        case GGML_TYPE_Q5_K:
+        case GGML_TYPE_Q6_K:
+            return true;
+        default:
+            return false;
+    }
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq2_xxs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int i   = blockIdx.x;
+    const block_iq2_xxs * x = (const block_iq2_xxs  *) vx;
+
+    const int tid = threadIdx.x;
+#if QK_K == 256
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * q2 = x[i].qs + 4*ib;
+    const uint8_t  * aux8 = (const uint8_t *)q2;
+    const uint8_t  * grid = (const uint8_t *)(iq2xxs_grid + aux8[il]);
+    const uint32_t aux32 = q2[2] | (q2[3] << 16);
+    const float d = (float)x[i].d * (0.5f + (aux32 >> 28)) * 0.25f;
+    const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*il) & 127];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+#else
+    assert(false);
+#endif
+
+}
+
+template<typename dst_t>
+static __global__ void dequantize_block_iq2_xs(const void * __restrict__ vx, dst_t * __restrict__ yy) {
+
+    const int i   = blockIdx.x;
+    const block_iq2_xs * x = (const block_iq2_xs *) vx;
+
+    const int tid = threadIdx.x;
+#if QK_K == 256
+    const int il = tid/8; // 0...3
+    const int ib = tid%8; // 0...7
+    dst_t * y = yy + i*QK_K + 32*ib + 8*il;
+    const uint16_t * q2 = x[i].qs + 4*ib;
+    const uint8_t  * grid = (const uint8_t *)(iq2xs_grid + (q2[il] & 511));
+    const float d = (float)x[i].d * (0.5f + ((x[i].scales[ib] >> 4*(il/2)) & 0xf)) * 0.25f;
+    const uint8_t signs = ksigns_iq2xs[q2[il] >> 9];
+    for (int j = 0; j < 8; ++j) y[j] = d * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+#else
+    assert(false);
+#endif
+
+}
+
 static __global__ void dequantize_mul_mat_vec_q2_k(const void * __restrict__ vx, const float * __restrict__ yy, float * __restrict__ dst, const int ncols, int nrows) {
 
     static_assert(16%K_QUANTS_PER_ITERATION == 0, "16 must be divisible by K_QUANTS_PER_ITERATION");
@@ -3825,6 +4145,91 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1_mul_mat(
     return vec_dot_q6_K_q8_1_impl_mmq(&x_ql[index_x], &y_qs[index_y], sc, x_dmf[i * (WARP_SIZE/QI6_K) + i/QI6_K], &y_df[index_y/QI8_1]);
 }
 
+static __device__ __forceinline__ float vec_dot_iq2_xxs_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if QK_K == 256
+    const block_iq2_xxs * bq2 = (const block_iq2_xxs *) vbq;
+
+#if QR2_XXS == 8
+    const int ib32 = iqs;
+    const uint16_t * q2 = bq2->qs + 4*ib32;
+    const uint8_t  * aux8 = (const uint8_t *)q2;
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    uint32_t aux32 = q2[2] | (q2[3] << 16);
+    int sumi = 0;
+    for (int l = 0; l < 4; ++l) {
+        const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+        const uint8_t  signs = ksigns_iq2xs[aux32 & 127];
+        for (int j = 0; j < 8; ++j) {
+            sumi += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+        }
+        q8 += 8;
+        aux32 >>= 7;
+    }
+    const float d = (float)bq2->d * (0.5f + aux32) * (float)bq8_1[ib32].ds.x * 0.25f;
+    return d * sumi;
+#else
+    // iqs is 0...15
+    const int ib32 = iqs/2;
+    const int il = iqs%2;
+    const uint16_t * q2 = bq2->qs + 4*ib32;
+    const uint8_t  * aux8 = (const uint8_t *)q2;
+    const uint8_t  * grid1 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
+    const uint8_t  * grid2 = (const uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
+    const uint32_t aux32 = q2[2] | (q2[3] << 16);
+    const float d = (float)bq2->d * (0.5f + (aux32 >> 28)) * (float)bq8_1[ib32].ds.x * 0.25f;
+    const uint8_t signs1 = ksigns_iq2xs[(aux32 >> 14*il) & 127];
+    const uint8_t signs2 = ksigns_iq2xs[(aux32 >> (14*il + 7)) & 127];
+    const int8_t * q8 = bq8_1[ib32].qs + 16*il;
+    int sumi1 = 0, sumi2 = 0;
+    for (int j = 0; j < 8; ++j) {
+        sumi1 += q8[j+0] * grid1[j] * (signs1 & kmask_iq2xs[j] ? -1 : 1);
+        sumi2 += q8[j+8] * grid2[j] * (signs2 & kmask_iq2xs[j] ? -1 : 1);
+    }
+    return d * (sumi1 + sumi2);
+#endif
+#else
+    assert(false);
+    return 0.f;
+#endif
+}
+
+static __device__ __forceinline__ float vec_dot_iq2_xs_q8_1(
+    const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs) {
+#if QK_K == 256
+    const block_iq2_xs * bq2 = (const block_iq2_xs *) vbq;
+
+    const int ib32 = iqs;
+    const uint16_t * q2 = bq2->qs + 4*ib32;
+    const int8_t   * q8 = bq8_1[ib32].qs;
+    const uint8_t ls1 = bq2->scales[ib32] & 0xf;
+    const uint8_t ls2 = bq2->scales[ib32] >>  4;
+    int sumi1 = 0;
+    for (int l = 0; l < 2; ++l) {
+        const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+        const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+        for (int j = 0; j < 8; ++j) {
+            sumi1 += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+        }
+        q8 += 8;
+    }
+    int sumi2 = 0;
+    for (int l = 2; l < 4; ++l) {
+        const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+        const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+        for (int j = 0; j < 8; ++j) {
+            sumi2 += q8[j] * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+        }
+        q8 += 8;
+    }
+    const float d = (float)bq2->d * (float)bq8_1[ib32].ds.x * 0.25f;
+    return d * ((0.5f + ls1) * sumi1 + (0.5f + ls2) * sumi2);
+#else
+    assert(false);
+    return 0.f;
+#endif
+}
+
 template <int qk, int qr, int qi, bool need_sum, typename block_q_t, int mmq_x, int mmq_y, int nwarps,
               allocate_tiles_cuda_t allocate_tiles, load_tiles_cuda_t load_tiles, int vdr, vec_dot_q_mul_mat_cuda_t vec_dot>
 static __device__ __forceinline__ void mul_mat_q(
@@ -5206,75 +5611,233 @@ static __global__ void diag_mask_inf_f32(const float * x, float * dst, const int
     dst[i] = x[i] - (col > n_past + row % rows_per_channel) * FLT_MAX;
 }
 
-static __global__ void soft_max_f32(const float * x, const float * y, float * dst, const int ncols, const int nrows_y, const float scale) {
+template <bool vals_smem, int ncols_template, int block_size_template, bool need_check>
+static __global__ void soft_max_f16(const float * x, const float * y, float * dst, const int ncols_par, const int nrows_y, const float scale) {
+#if !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+    const int ncols_data = ncols_template == 0 ? ncols_par : ncols_template;
+    const int ncols_smem = GGML_PAD(ncols_data, 2*WARP_SIZE)/2;
+
     const int tid  = threadIdx.x;
     const int rowx = blockIdx.x;
     const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension
 
-    const int block_size = blockDim.x;
+    const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
 
     const int warp_id = threadIdx.x / WARP_SIZE;
     const int lane_id = threadIdx.x % WARP_SIZE;
 
-    __shared__ float buf[CUDA_SOFT_MAX_BLOCK_SIZE/WARP_SIZE];
+    extern __shared__ half data_soft_max_f16[];
+    half  * buf_iw = data_soft_max_f16 + 0; // shared memory buffer for inter-warp communication
+    // (shared memory) buffer to cache values between iterations:
+    half2 * vals = vals_smem ? (half2 *) (buf_iw + WARP_SIZE) : (half2 *) (dst + rowx*ncols_data);
+    // if the buffer is larger than max. shared memory per block, use dst as temp. buffer instead
+    // in that case col_smem == col_data must be enforced to avoid race conditions
 
-    float max_val = -INFINITY;
+    half2 max_val = make_half2(-INFINITY, -INFINITY);
 
-    for (int col = tid; col < ncols; col += block_size) {
-        const int ix = rowx*ncols + col;
-        const int iy = rowy*ncols + col;
-        max_val = max(max_val, x[ix]*scale + (y ? y[iy] : 0.0f));
+#pragma unroll
+    for (int col0 = 0; col0 < ncols_smem; col0 += block_size) {
+        const int col_data = 2*col0 + 2*WARP_SIZE*warp_id + lane_id;
+        const int col_smem = vals_smem ? col0 + tid : col_data;
+
+        const int ix = rowx*ncols_data + col_data;
+        const int iy = rowy*ncols_data + col_data;
+
+        half2 val;
+        if (need_check && col_data + 0 >= ncols_data) {
+            val.x = -INFINITY;
+        } else {
+            val.x = x[ix + 0]*scale + (y ? y[iy + 0] : 0.0f);
+        }
+        if (need_check && col_data + WARP_SIZE >= ncols_data) {
+            val.y = -INFINITY;
+        } else {
+            val.y = x[ix + WARP_SIZE]*scale + (y ? y[iy + WARP_SIZE] : 0.0f);
+        }
+        if (!need_check || col_smem < (vals_smem ? ncols_smem : ncols_data)) {
+            vals[col_smem] = val;
+        }
+        max_val = __hmax2(max_val, val);
     }
 
     // find the max value in the block
     max_val = warp_reduce_max(max_val);
     if (block_size > WARP_SIZE) {
         if (warp_id == 0) {
-            buf[lane_id] = -INFINITY;
+            buf_iw[lane_id] = -INFINITY;
         }
         __syncthreads();
 
         if (lane_id == 0) {
-            buf[warp_id] = max_val;
+            buf_iw[warp_id] = __hmax(max_val.x, max_val.y);
         }
         __syncthreads();
 
-        max_val = buf[lane_id];
+        max_val = __half2half2(buf_iw[lane_id]);
         max_val = warp_reduce_max(max_val);
+    } else {
+        max_val = __half2half2(__hmax(max_val.x, max_val.y));
     }
 
-    float tmp = 0.f;
+    half2 tmp = make_half2(0.0f, 0.0f); // partial sums
+
+#pragma unroll
+    for (int col0 = 0; col0 < ncols_smem; col0 += block_size) {
+        const int col_smem = vals_smem ? col0 + tid : 2*col0 + 2*warp_id*WARP_SIZE + lane_id;
+
+        if (ncols_template == 0 && col_smem >= (vals_smem ? ncols_smem : ncols_data)) {
+            break;
+        }
+
+        const half2 val = h2exp(vals[col_smem] - max_val);
 
-    for (int col = tid; col < ncols; col += block_size) {
-        const int ix = rowx*ncols + col;
-        const int iy = rowy*ncols + col;
-        const float val = expf((x[ix]*scale + (y ? y[iy] : 0.0f)) - max_val);
         tmp += val;
-        dst[ix] = val;
+        vals[col_smem] = val;
     }
 
     // find the sum of exps in the block
     tmp = warp_reduce_sum(tmp);
     if (block_size > WARP_SIZE) {
         if (warp_id == 0) {
-            buf[lane_id] = 0.f;
+            buf_iw[lane_id] = 0.0f;
         }
         __syncthreads();
 
         if (lane_id == 0) {
-            buf[warp_id] = tmp;
+            buf_iw[warp_id] = tmp.x + tmp.y;
         }
         __syncthreads();
 
-        tmp = buf[lane_id];
+        tmp = __half2half2(buf_iw[lane_id]);
+        tmp = warp_reduce_sum(tmp);
+    } else {
+        tmp = __half2half2(tmp.x + tmp.y);
+    }
+
+    const half2 inv_sum = make_half2(1.0f, 1.0f) / tmp;
+
+#pragma unroll
+    for (int col0 = 0; col0 < ncols_smem; col0 += block_size) {
+        const int col_data = 2*col0 + 2*WARP_SIZE*warp_id + lane_id;
+        const int col_smem = vals_smem ? col0 + tid : col_data;
+
+        const int idst = rowx*ncols_data + col_data;
+        const half2 result = vals[col_smem] * inv_sum;
+
+        if (need_check && col_data + 0 >= ncols_data) {
+            return;
+        }
+        dst[idst] = result.x;
+
+        if (need_check && col_data + WARP_SIZE >= ncols_data) {
+            return;
+        }
+
+        dst[idst + WARP_SIZE] = result.y;
+    }
+#else
+    (void) x; (void) y; (void) dst; (void) ncols_par; (void) nrows_y; (void) scale;
+    bad_arch();
+#endif // !(defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)) && __CUDA_ARCH__ >= CC_PASCAL
+}
+
+template <bool vals_smem, int ncols_template, int block_size_template>
+static __global__ void soft_max_f32(const float * x, const float * y, float * dst, const int ncols_par, const int nrows_y, const float scale) {
+    const int ncols = ncols_template == 0 ? ncols_par : ncols_template;
+
+    const int tid  = threadIdx.x;
+    const int rowx = blockIdx.x;
+    const int rowy = rowx % nrows_y; // broadcast the mask (y) in the row dimension
+
+    const int block_size = block_size_template == 0 ? blockDim.x : block_size_template;
+
+    const int warp_id = threadIdx.x / WARP_SIZE;
+    const int lane_id = threadIdx.x % WARP_SIZE;
+
+    extern __shared__ float data_soft_max_f32[];
+    float * buf_iw = data_soft_max_f32; // shared memory buffer for inter-warp communication
+    // shared memory buffer to cache values between iterations:
+    float * vals = vals_smem ? buf_iw + WARP_SIZE : dst + rowx*ncols;
+
+    float max_val = -INFINITY;
+
+#pragma unroll
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+
+        if (ncols_template == 0 && col >= ncols) {
+            break;
+        }
+
+        const int ix = rowx*ncols + col;
+        const int iy = rowy*ncols + col;
+
+        const float val = x[ix]*scale + (y ? y[iy] : 0.0f);
+        vals[col] = val;
+        max_val = max(max_val, val);
+    }
+
+    // find the max value in the block
+    max_val = warp_reduce_max(max_val);
+    if (block_size > WARP_SIZE) {
+        if (warp_id == 0) {
+            buf_iw[lane_id] = -INFINITY;
+        }
+        __syncthreads();
+
+        if (lane_id == 0) {
+            buf_iw[warp_id] = max_val;
+        }
+        __syncthreads();
+
+        max_val = buf_iw[lane_id];
+        max_val = warp_reduce_max(max_val);
+    }
+
+    float tmp = 0.0f; // partial sum
+
+#pragma unroll
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+
+        if (ncols_template == 0 && col >= ncols) {
+            break;
+        }
+
+        const float val = expf(vals[col] - max_val);
+        tmp += val;
+        vals[col] = val;
+    }
+
+    // find the sum of exps in the block
+    tmp = warp_reduce_sum(tmp);
+    if (block_size > WARP_SIZE) {
+        if (warp_id == 0) {
+            buf_iw[lane_id] = 0.0f;
+        }
+        __syncthreads();
+
+        if (lane_id == 0) {
+            buf_iw[warp_id] = tmp;
+        }
+        __syncthreads();
+
+        tmp = buf_iw[lane_id];
         tmp = warp_reduce_sum(tmp);
     }
 
-    const float inv_tmp = 1.f / tmp;
+    const float inv_sum = 1.0f / tmp;
 
-    for (int col = tid; col < ncols; col += block_size) {
-        const int i = rowx*ncols + col;
-        dst[i] *= inv_tmp;
+#pragma unroll
+    for (int col0 = 0; col0 < ncols; col0 += block_size) {
+        const int col = col0 + tid;
+
+        if (ncols_template == 0 && col >= ncols) {
+            return;
+        }
+
+        const int idst = rowx*ncols + col;
+        dst[idst] = vals[col] * inv_sum;
     }
 }
 
@@ -5664,6 +6227,18 @@ static void dequantize_row_q6_K_cuda(const void * vx, dst_t * y, const int k, cu
 #endif
 }
 
+template<typename dst_t>
+static void dequantize_row_iq2_xxs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq2_xxs<<<nb, 32, 0, stream>>>(vx, y);
+}
+
+template<typename dst_t>
+static void dequantize_row_iq2_xs_cuda(const void * vx, dst_t * y, const int k, cudaStream_t stream) {
+    const int nb = k / QK_K;
+    dequantize_block_iq2_xs<<<nb, 32, 0, stream>>>(vx, y);
+}
+
 template <typename src_t, typename dst_t>
 static void convert_unary_cuda(const void * __restrict__ vx, dst_t * __restrict__ y, const int k, cudaStream_t stream) {
     const int num_blocks = (k + CUDA_DEQUANTIZE_BLOCK_SIZE - 1) / CUDA_DEQUANTIZE_BLOCK_SIZE;
@@ -5692,6 +6267,10 @@ static to_fp16_cuda_t ggml_get_to_fp16_cuda(ggml_type type) {
             return dequantize_row_q5_K_cuda;
         case GGML_TYPE_Q6_K:
             return dequantize_row_q6_K_cuda;
+        case GGML_TYPE_IQ2_XXS:
+            return dequantize_row_iq2_xxs_cuda;
+        case GGML_TYPE_IQ2_XS:
+            return dequantize_row_iq2_xs_cuda;
         case GGML_TYPE_F32:
             return convert_unary_cuda<float>;
         default:
@@ -5721,6 +6300,10 @@ static to_fp32_cuda_t ggml_get_to_fp32_cuda(ggml_type type) {
             return dequantize_row_q5_K_cuda;
         case GGML_TYPE_Q6_K:
             return dequantize_row_q6_K_cuda;
+        case GGML_TYPE_IQ2_XXS:
+            return dequantize_row_iq2_xxs_cuda;
+        case GGML_TYPE_IQ2_XS:
+            return dequantize_row_iq2_xs_cuda;
         case GGML_TYPE_F16:
             return convert_unary_cuda<half>;
         default:
@@ -5915,6 +6498,24 @@ static void mul_mat_vec_q6_K_q8_1_cuda(const void * vx, const void * vy, float *
         <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
 }
 
+static void mul_mat_vec_iq2_xxs_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    mul_mat_vec_q<QK_K, QI2_XXS, block_iq2_xxs, 1, vec_dot_iq2_xxs_q8_1>
+        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
+}
+
+static void mul_mat_vec_iq2_xs_q8_1_cuda(const void * vx, const void * vy, float * dst, const int ncols, const int nrows, cudaStream_t stream) {
+    GGML_ASSERT(ncols % QK_K == 0);
+    const int block_num_y = (nrows + GGML_CUDA_MMV_Y - 1) / GGML_CUDA_MMV_Y;
+    const dim3 block_nums(block_num_y, 1, 1);
+    const dim3 block_dims(WARP_SIZE, GGML_CUDA_MMV_Y, 1);
+    mul_mat_vec_q<QK_K, QI2_XS, block_iq2_xs, 1, vec_dot_iq2_xs_q8_1>
+        <<<block_nums, block_dims, 0, stream>>>(vx, vy, dst, ncols, nrows);
+}
+
 static void ggml_mul_mat_q4_0_q8_1_cuda(
     const void * vx, const void * vy, float * dst, const int ncols_x, const int nrows_x,
     const int ncols_y, const int nrows_y, const int nrows_dst, cudaStream_t stream) {
@@ -6554,12 +7155,90 @@ static void diag_mask_inf_f32_cuda(const float * x, float * dst, const int ncols
     diag_mask_inf_f32<<<block_nums, block_dims, 0, stream>>>(x, dst, ncols_x, rows_per_channel, n_past);
 }
 
+static void soft_max_f16_cuda(const float * x, const float * y, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, cudaStream_t stream) {
+    int nth = WARP_SIZE;
+    while (nth < ncols_x/2 && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
+    const dim3 block_dims(nth,     1, 1);
+    const dim3 block_nums(nrows_x, 1, 1);
+    const size_t shmem = (GGML_PAD(ncols_x, 2*WARP_SIZE) + WARP_SIZE)*sizeof(half);
+    static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
+    if (shmem <= g_device_caps[g_main_device].smpb) {
+        switch (ncols_x) {
+            case 32:
+                soft_max_f16<true, 32, 32, true><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 64:
+                soft_max_f16<true, 64, 32, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 128:
+                soft_max_f16<true, 128, 64, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 256:
+                soft_max_f16<true, 256, 128, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 512:
+                soft_max_f16<true, 512, 256, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 1024:
+                soft_max_f16<true, 1024, 512, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 2048:
+                soft_max_f16<true, 2048, 1024, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 4096:
+                soft_max_f16<true, 4096, 1024, false><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            default:
+                soft_max_f16<true, 0, 0, true><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+        }
+    } else {
+        const size_t shmem_low = WARP_SIZE*sizeof(half);
+        soft_max_f16<false, 0, 0, true><<<block_nums, block_dims, shmem_low, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+    }
+}
+
 static void soft_max_f32_cuda(const float * x, const float * y, float * dst, const int ncols_x, const int nrows_x, const int nrows_y, const float scale, cudaStream_t stream) {
     int nth = WARP_SIZE;
     while (nth < ncols_x && nth < CUDA_SOFT_MAX_BLOCK_SIZE) nth *= 2;
     const dim3 block_dims(nth,     1, 1);
     const dim3 block_nums(nrows_x, 1, 1);
-    soft_max_f32<<<block_nums, block_dims, 0, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+    const size_t shmem = (GGML_PAD(ncols_x, WARP_SIZE) + WARP_SIZE)*sizeof(float);
+    static_assert(CUDA_SOFT_MAX_BLOCK_SIZE == 1024, "These values need to be adjusted.");
+    if (shmem < g_device_caps[g_main_device].smpb) {
+        switch (ncols_x) {
+            case 32:
+                soft_max_f32<true, 32, 32><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 64:
+                soft_max_f32<true, 64, 64><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 128:
+                soft_max_f32<true, 128, 128><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 256:
+                soft_max_f32<true, 256, 256><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 512:
+                soft_max_f32<true, 512, 512><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 1024:
+                soft_max_f32<true, 1024, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 2048:
+                soft_max_f32<true, 2048, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            case 4096:
+                soft_max_f32<true, 4096, 1024><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+            default:
+                soft_max_f32<true, 0, 0><<<block_nums, block_dims, shmem, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+                break;
+        }
+    } else {
+        const size_t shmem_low = WARP_SIZE*sizeof(float);
+        soft_max_f32<false, 0, 0><<<block_nums, block_dims, shmem_low, stream>>>(x, y, dst, ncols_x, nrows_y, scale);
+    }
 }
 
 static void im2col_f32_f16_cuda(const float* x, half* dst,
@@ -6874,6 +7553,7 @@ void ggml_init_cublas() {
 #else
             g_device_caps[id].cc = 100*prop.major + 10*prop.minor;
 #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+            g_device_caps[id].smpb = prop.sharedMemPerBlock;
         }
         for (int id = 0; id < g_device_count; ++id) {
             g_tensor_split[id] /= total_vram;
@@ -7407,6 +8087,8 @@ static int64_t get_row_rounding(ggml_type type) {
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
             return max_compute_capability >= CC_RDNA2 ? 128 : 64;
         default:
             GGML_ASSERT(false);
@@ -7427,6 +8109,8 @@ static int64_t get_row_rounding(ggml_type type) {
         case GGML_TYPE_Q3_K:
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
             return max_compute_capability >= CC_VOLTA ? 128 : 64;
         case GGML_TYPE_Q6_K:
             return 64;
@@ -7477,6 +8161,12 @@ static void ggml_cuda_op_mul_mat_vec_q(
         case GGML_TYPE_Q6_K:
             mul_mat_vec_q6_K_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
             break;
+        case GGML_TYPE_IQ2_XXS:
+            mul_mat_vec_iq2_xxs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
+        case GGML_TYPE_IQ2_XS:
+            mul_mat_vec_iq2_xs_q8_1_cuda(src0_dd_i, src1_ddq_i, dst_dd_i, ne00, row_diff, stream);
+            break;
         default:
             GGML_ASSERT(false);
             break;
@@ -7884,7 +8574,21 @@ static void ggml_cuda_op_soft_max(
     float scale = 1.0f;
     memcpy(&scale, dst->op_params, sizeof(float));
 
-    soft_max_f32_cuda(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, main_stream);
+#if defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+    const bool use_f16_soft_max = false;
+#else
+#ifdef GGML_CUDA_F16
+    const bool use_f16_soft_max = true;
+#else
+    const bool use_f16_soft_max = false;
+#endif // GGML_CUDA_F16
+#endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
+
+    if (use_f16_soft_max) {
+        soft_max_f16_cuda(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, main_stream);
+    } else {
+        soft_max_f32_cuda(src0_dd, src1 ? src1_dd : nullptr, dst_dd, ne00, nrows_x, nrows_y, scale, main_stream);
+    }
 
     (void) dst;
 }
@@ -8693,6 +9397,8 @@ static void ggml_cuda_mul_mat(const ggml_tensor * src0, const ggml_tensor * src1
 
 #endif // defined(GGML_USE_HIPBLAS) && defined(__HIP_PLATFORM_AMD__)
 
+    use_mul_mat_q = use_mul_mat_q && ggml_cuda_supports_mmq(src0->type);
+
     // debug helpers
     //printf("src0: %8d %8d %8d %8d\n", src0->ne[0], src0->ne[1], src0->ne[2], src0->ne[3]);
     //printf("      %8d %8d %8d %8d\n", src0->nb[0], src0->nb[1], src0->nb[2], src0->nb[3]);
@@ -9700,8 +10406,8 @@ static void ggml_backend_cuda_buffer_set_tensor(ggml_backend_buffer_t buffer, gg
 
     ggml_cuda_set_device(ctx->device);
     CUDA_CHECK(cudaDeviceSynchronize());
-
     CUDA_CHECK(cudaMemcpy((char *)tensor->data + offset, data, size, cudaMemcpyHostToDevice));
+    CUDA_CHECK(cudaDeviceSynchronize());
 }
 
 static void ggml_backend_cuda_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {

ggml-metal.m

@@ -88,6 +88,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(get_rows_q5_K);
     GGML_METAL_DECL_KERNEL(get_rows_q6_K);
     GGML_METAL_DECL_KERNEL(get_rows_i32);
+    GGML_METAL_DECL_KERNEL(get_rows_iq2_xxs);
+    GGML_METAL_DECL_KERNEL(get_rows_iq2_xs);
     GGML_METAL_DECL_KERNEL(rms_norm);
     GGML_METAL_DECL_KERNEL(group_norm);
     GGML_METAL_DECL_KERNEL(norm);
@@ -106,6 +108,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mv_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_iq2_xxs_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_iq2_xs_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_id_f32_f32);
     //GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f16);
     GGML_METAL_DECL_KERNEL(mul_mv_id_f16_f32);
@@ -121,6 +125,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mv_id_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_id_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mv_id_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_iq2_xxs_f32);
+    GGML_METAL_DECL_KERNEL(mul_mv_id_iq2_xs_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_f32_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q4_0_f32);
@@ -133,6 +139,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mm_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_iq2_xxs_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_iq2_xs_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_id_f32_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_id_f16_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_id_q4_0_f32);
@@ -145,6 +153,8 @@ struct ggml_metal_context {
     GGML_METAL_DECL_KERNEL(mul_mm_id_q4_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_id_q5_K_f32);
     GGML_METAL_DECL_KERNEL(mul_mm_id_q6_K_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_iq2_xxs_f32);
+    GGML_METAL_DECL_KERNEL(mul_mm_id_iq2_xs_f32);
     GGML_METAL_DECL_KERNEL(rope_f32);
     GGML_METAL_DECL_KERNEL(rope_f16);
     GGML_METAL_DECL_KERNEL(alibi_f32);
@@ -379,6 +389,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(get_rows_q5_K);
         GGML_METAL_ADD_KERNEL(get_rows_q6_K);
         GGML_METAL_ADD_KERNEL(get_rows_i32);
+        GGML_METAL_ADD_KERNEL(get_rows_iq2_xxs);
+        GGML_METAL_ADD_KERNEL(get_rows_iq2_xs);
         GGML_METAL_ADD_KERNEL(rms_norm);
         GGML_METAL_ADD_KERNEL(group_norm);
         GGML_METAL_ADD_KERNEL(norm);
@@ -397,6 +409,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul_mv_q4_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_q5_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_q6_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_iq2_xxs_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_iq2_xs_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_id_f32_f32);
         //GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f16);
         GGML_METAL_ADD_KERNEL(mul_mv_id_f16_f32);
@@ -412,6 +426,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
         GGML_METAL_ADD_KERNEL(mul_mv_id_q4_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_id_q5_K_f32);
         GGML_METAL_ADD_KERNEL(mul_mv_id_q6_K_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_iq2_xxs_f32);
+        GGML_METAL_ADD_KERNEL(mul_mv_id_iq2_xs_f32);
         if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) {
             GGML_METAL_ADD_KERNEL(mul_mm_f32_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_f16_f32);
@@ -425,6 +441,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
             GGML_METAL_ADD_KERNEL(mul_mm_q4_K_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_q5_K_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_q6_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_iq2_xxs_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_iq2_xs_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_id_f32_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_id_f16_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_id_q4_0_f32);
@@ -437,6 +455,8 @@ struct ggml_metal_context * ggml_metal_init(int n_cb) {
             GGML_METAL_ADD_KERNEL(mul_mm_id_q4_K_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_id_q5_K_f32);
             GGML_METAL_ADD_KERNEL(mul_mm_id_q6_K_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_iq2_xxs_f32);
+            GGML_METAL_ADD_KERNEL(mul_mm_id_iq2_xs_f32);
         }
         GGML_METAL_ADD_KERNEL(rope_f32);
         GGML_METAL_ADD_KERNEL(rope_f16);
@@ -502,6 +522,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(get_rows_q5_K);
     GGML_METAL_DEL_KERNEL(get_rows_q6_K);
     GGML_METAL_DEL_KERNEL(get_rows_i32);
+    GGML_METAL_DEL_KERNEL(get_rows_iq2_xxs);
+    GGML_METAL_DEL_KERNEL(get_rows_iq2_xs);
     GGML_METAL_DEL_KERNEL(rms_norm);
     GGML_METAL_DEL_KERNEL(group_norm);
     GGML_METAL_DEL_KERNEL(norm);
@@ -520,6 +542,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(mul_mv_q4_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_q5_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_q6_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_iq2_xxs_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_iq2_xs_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_id_f32_f32);
     //GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f16);
     GGML_METAL_DEL_KERNEL(mul_mv_id_f16_f32);
@@ -535,6 +559,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
     GGML_METAL_DEL_KERNEL(mul_mv_id_q4_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_id_q5_K_f32);
     GGML_METAL_DEL_KERNEL(mul_mv_id_q6_K_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_iq2_xxs_f32);
+    GGML_METAL_DEL_KERNEL(mul_mv_id_iq2_xs_f32);
     if ([ctx->device supportsFamily:MTLGPUFamilyApple7]) {
         GGML_METAL_DEL_KERNEL(mul_mm_f32_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_f16_f32);
@@ -548,6 +574,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
         GGML_METAL_DEL_KERNEL(mul_mm_q4_K_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_q5_K_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_q6_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_iq2_xxs_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_iq2_xs_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_id_f32_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_id_f16_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_id_q4_0_f32);
@@ -560,6 +588,8 @@ void ggml_metal_free(struct ggml_metal_context * ctx) {
         GGML_METAL_DEL_KERNEL(mul_mm_id_q4_K_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_id_q5_K_f32);
         GGML_METAL_DEL_KERNEL(mul_mm_id_q6_K_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_iq2_xxs_f32);
+        GGML_METAL_DEL_KERNEL(mul_mm_id_iq2_xs_f32);
     }
     GGML_METAL_DEL_KERNEL(rope_f32);
     GGML_METAL_DEL_KERNEL(rope_f16);
@@ -1052,6 +1082,10 @@ bool ggml_metal_graph_compute(
                     GGML_ASSERT(!"unsupported op");
                 }
 
+#ifndef GGML_METAL_NDEBUG
+                [encoder pushDebugGroup:[NSString stringWithCString:ggml_op_desc(dst) encoding:NSUTF8StringEncoding]];
+#endif
+
                 const int64_t  ne00 = src0 ? src0->ne[0] : 0;
                 const int64_t  ne01 = src0 ? src0->ne[1] : 0;
                 const int64_t  ne02 = src0 ? src0->ne[2] : 0;
@@ -1541,6 +1575,8 @@ bool ggml_metal_graph_compute(
                                     case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q4_K_f32]; break;
                                     case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q5_K_f32]; break;
                                     case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_q6_K_f32]; break;
+                                    case GGML_TYPE_IQ2_XXS: [encoder setComputePipelineState:ctx->pipeline_mul_mm_iq2_xxs_f32]; break;
+                                    case GGML_TYPE_IQ2_XS : [encoder setComputePipelineState:ctx->pipeline_mul_mm_iq2_xs_f32]; break;
                                     default: GGML_ASSERT(false && "MUL MAT-MAT not implemented");
                                 }
                                 [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
@@ -1653,6 +1689,18 @@ bool ggml_metal_graph_compute(
                                             nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_q6_K_f32];
                                         } break;
+                                    case GGML_TYPE_IQ2_XXS:
+                                        {
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_iq2_xxs_f32];
+                                        } break;
+                                    case GGML_TYPE_IQ2_XS:
+                                        {
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_iq2_xs_f32];
+                                        } break;
                                     default:
                                         {
                                             GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src0t);
@@ -1689,6 +1737,11 @@ bool ggml_metal_graph_compute(
                                     src0t == GGML_TYPE_Q2_K) { // || src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
+                                else if (src0t == GGML_TYPE_IQ2_XXS || src0t == GGML_TYPE_IQ2_XS) {
+                                    const int mem_size = src0t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
+                                    [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 7)/8, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
                                 else if (src0t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne01 + 3)/4, ne11, ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
@@ -1778,6 +1831,8 @@ bool ggml_metal_graph_compute(
                                     case GGML_TYPE_Q4_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q4_K_f32]; break;
                                     case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q5_K_f32]; break;
                                     case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_q6_K_f32]; break;
+                                    case GGML_TYPE_IQ2_XXS: [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_iq2_xxs_f32]; break;
+                                    case GGML_TYPE_IQ2_XS : [encoder setComputePipelineState:ctx->pipeline_mul_mm_id_iq2_xs_f32]; break;
                                     default: GGML_ASSERT(false && "MUL_MAT_ID not implemented");
                                 }
                                 [encoder setBuffer:id_src0 offset:offs_src0    atIndex:0];
@@ -1893,6 +1948,18 @@ bool ggml_metal_graph_compute(
                                             nth1 = 32;
                                             [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_q6_K_f32];
                                         } break;
+                                    case GGML_TYPE_IQ2_XXS:
+                                        {
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_iq2_xxs_f32];
+                                        } break;
+                                    case GGML_TYPE_IQ2_XS:
+                                        {
+                                            nth0 = 4;
+                                            nth1 = 16;
+                                            [encoder setComputePipelineState:ctx->pipeline_mul_mv_id_iq2_xs_f32];
+                                        } break;
                                     default:
                                         {
                                             GGML_METAL_LOG_ERROR("Asserting on type %d\n", (int)src2t);
@@ -1945,6 +2012,11 @@ bool ggml_metal_graph_compute(
                                     src2t == GGML_TYPE_Q2_K) { // || src2t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
+                                else if (src2t == GGML_TYPE_IQ2_XXS || src2t == GGML_TYPE_IQ2_XS) {
+                                    const int mem_size = src2t == GGML_TYPE_IQ2_XXS ? 256*8+128 : 512*8+128;
+                                    [encoder setThreadgroupMemoryLength:mem_size atIndex:0];
+                                    [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 7)/8, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
+                                }
                                 else if (src2t == GGML_TYPE_Q4_K) {
                                     [encoder dispatchThreadgroups:MTLSizeMake((ne21 + 3)/4, _ne1, ne01*ne12*ne13) threadsPerThreadgroup:MTLSizeMake(nth0, nth1, 1)];
                                 }
@@ -1982,6 +2054,8 @@ bool ggml_metal_graph_compute(
                                 case GGML_TYPE_Q5_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q5_K]; break;
                                 case GGML_TYPE_Q6_K: [encoder setComputePipelineState:ctx->pipeline_get_rows_q6_K]; break;
                                 case GGML_TYPE_I32:  [encoder setComputePipelineState:ctx->pipeline_get_rows_i32];  break;
+                                case GGML_TYPE_IQ2_XXS: [encoder setComputePipelineState:ctx->pipeline_get_rows_iq2_xxs]; break;
+                                case GGML_TYPE_IQ2_XS : [encoder setComputePipelineState:ctx->pipeline_get_rows_iq2_xs]; break;
                                 default: GGML_ASSERT(false && "not implemented");
                             }
 
@@ -2383,6 +2457,10 @@ bool ggml_metal_graph_compute(
                             GGML_ASSERT(false);
                         }
                 }
+
+#ifndef GGML_METAL_NDEBUG
+                [encoder popDebugGroup];
+#endif
             }
 
             if (encoder != nil) {

ggml-metal.metal

@@ -2446,6 +2446,19 @@ typedef struct {
 } block_q6_K;
 // 210 bytes / block
 
+typedef struct {
+    half d;
+    uint16_t qs[QK_K/8];
+} block_iq2_xxs;
+// 66 bytes / block for QK_K = 256, so 2.0625 bpw
+
+typedef struct {
+    half d;
+    uint16_t qs[QK_K/8];
+    uint8_t  scales[QK_K/32];
+} block_iq2_xs;
+// 74 bytes / block for QK_K = 256, so 2.3125 bpw
+
 //====================================== dot products =========================
 
 void kernel_mul_mv_q2_K_f32_impl(
@@ -3468,6 +3481,495 @@ kernel void kernel_mul_mv_q6_K_f32(
     kernel_mul_mv_q6_K_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, tgpig, tiisg, sgitg);
 }
 
+// ======================= "True" 2-bit
+
+constexpr constant static uint64_t iq2xxs_grid[256] = {
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808,
+    0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819,
+    0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819,
+    0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b,
+    0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808,
+    0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08,
+    0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b,
+    0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819,
+    0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08,
+    0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808,
+    0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08,
+    0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808,
+    0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808,
+    0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919,
+    0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819,
+    0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08,
+    0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908,
+    0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819,
+    0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808,
+    0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808,
+    0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908,
+    0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808,
+    0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08,
+    0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819,
+    0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819,
+    0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819,
+    0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908,
+    0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19,
+    0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819,
+    0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b,
+    0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808,
+    0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908,
+    0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08,
+    0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08,
+    0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908,
+    0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819,
+    0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808,
+    0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808,
+    0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19,
+    0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819,
+    0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919,
+    0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b,
+    0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08,
+    0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808,
+    0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908,
+    0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b,
+    0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819,
+    0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08,
+    0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08,
+    0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808,
+    0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b,
+    0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b,
+    0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908,
+    0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819,
+    0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808,
+    0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908,
+    0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b,
+    0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808,
+    0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b,
+    0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b,
+    0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808,
+    0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19,
+    0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908,
+};
+
+constexpr constant static uint64_t iq2xs_grid[512] = {
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
+    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
+    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
+    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
+    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808,
+    0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819,
+    0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819,
+    0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808,
+    0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b,
+    0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b,
+    0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908,
+    0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908,
+    0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919,
+    0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808,
+    0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919,
+    0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908,
+    0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b,
+    0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908,
+    0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08,
+    0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808,
+    0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808,
+    0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819,
+    0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908,
+    0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819,
+    0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808,
+    0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b,
+    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819,
+    0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819,
+    0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808,
+    0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908,
+    0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19,
+    0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b,
+    0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b,
+    0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919,
+    0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808,
+    0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819,
+    0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819,
+    0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b,
+    0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908,
+    0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808,
+    0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819,
+    0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808,
+    0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919,
+    0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808,
+    0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808,
+    0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908,
+    0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908,
+    0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808,
+    0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b,
+    0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819,
+    0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919,
+    0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908,
+    0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808,
+    0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908,
+    0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919,
+    0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08,
+    0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19,
+    0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b,
+    0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b,
+    0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808,
+    0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08,
+    0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b,
+    0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908,
+    0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b,
+    0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908,
+    0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08,
+    0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808,
+    0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808,
+    0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08,
+    0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819,
+    0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919,
+    0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808,
+    0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808,
+    0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819,
+    0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819,
+    0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908,
+    0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908,
+    0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b,
+    0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908,
+    0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908,
+    0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908,
+    0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808,
+    0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819,
+    0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819,
+    0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819,
+    0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808,
+    0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b,
+    0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819,
+    0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819,
+    0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08,
+    0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808,
+    0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19,
+    0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919,
+    0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808,
+    0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19,
+    0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b,
+    0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808,
+    0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b,
+    0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b,
+    0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08,
+    0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b,
+    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808,
+    0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819,
+    0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808,
+    0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808,
+    0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08,
+    0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b,
+    0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19,
+    0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08,
+    0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919,
+    0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08,
+    0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08,
+    0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908,
+    0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908,
+    0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b,
+    0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908,
+    0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808,
+    0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b,
+    0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808,
+    0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808,
+    0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19,
+    0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08,
+    0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808,
+    0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b,
+    0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808,
+    0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b,
+    0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
+};
+
+constexpr constant static uint8_t ksigns_iq2xs[128] = {
+      0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
+    144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
+    160,  33,  34, 163,  36, 165, 166,  39,  40, 169, 170,  43, 172,  45,  46, 175,
+     48, 177, 178,  51, 180,  53,  54, 183, 184,  57,  58, 187,  60, 189, 190,  63,
+    192,  65,  66, 195,  68, 197, 198,  71,  72, 201, 202,  75, 204,  77,  78, 207,
+     80, 209, 210,  83, 212,  85,  86, 215, 216,  89,  90, 219,  92, 221, 222,  95,
+     96, 225, 226,  99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111,
+    240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
+};
+
+constexpr constant static uint8_t kmask_iq2xs[8] = {1, 2, 4, 8, 16, 32, 64, 128};
+
+void kernel_mul_mv_iq2_xxs_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_iq2_xxs * x = (device const block_iq2_xxs *) src0 + ib_row + offset0;
+    device const float         * y = (device const float         *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int nb32 = nb * (QK_K / 32);
+
+    threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
+    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 256);
+    {
+        int nval = 4;
+        int pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) values[pos + i] = iq2xxs_grid[pos + i];
+        nval = 2;
+        pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+#if QK_K == 256
+    const int ix = tiisg;
+
+    device const float * y4 = y + 32 * ix;
+
+    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
+
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
+
+        device const block_iq2_xxs * xr = x + ibl;
+        device const uint16_t * q2 = xr->qs + 4 * ib;
+        device const half * dh = &xr->d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            const float db = dh[0];
+            device const uint8_t * aux8 = (device const uint8_t *)q2;
+            const uint32_t aux32 = q2[2] | (q2[3] << 16);
+            const float d = db * (0.5f + (aux32 >> 28));
+
+            float sum = 0;
+            for (int l = 0; l < 4; ++l) {
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + aux8[l]);
+                const uint8_t signs = shared_signs[(aux32 >> 7*l) & 127];
+                for (int j = 0; j < 8; ++j) {
+                    sum += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+                }
+            }
+            sumf[row] += d * sum;
+
+            dh += nb*sizeof(block_iq2_xxs)/2;
+            q2 += nb*sizeof(block_iq2_xxs)/2;
+        }
+
+        y4 += 32 * 32;
+    }
+#else
+    // TODO
+#endif
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_iq2_xxs_f32")]]
+kernel void kernel_mul_mv_iq2_xxs_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq2_xxs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
+
+void kernel_mul_mv_iq2_xs_f32_impl(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    const int nb = ne00/QK_K;
+    const int r0 = tgpig.x;
+    const int r1 = tgpig.y;
+    const int im = tgpig.z;
+
+    const int first_row = (r0 * N_SIMDGROUP + sgitg) * N_DST;
+    const int ib_row = first_row * nb;
+
+    const uint i12 = im%ne12;
+    const uint i13 = im/ne12;
+
+    const uint offset0 = (i12/r2)*(nb*ne01) + (i13/r3)*(nb*ne01*ne02);
+
+    device const block_iq2_xs * x = (device const block_iq2_xs *) src0 + ib_row + offset0;
+    device const float        * y = (device const float        *) src1 + r1*ne10 + im*ne00*ne1;
+
+    float yl[32];
+    float sumf[N_DST]={0.f}, all_sum;
+
+    const int nb32 = nb * (QK_K / 32);
+
+    threadgroup uint64_t * values = (threadgroup uint64_t *)shared_values;
+    threadgroup uint8_t  * shared_signs = (threadgroup uint8_t *)(values + 512);
+    {
+        int nval = 8;
+        int pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) values[pos + i] = iq2xs_grid[pos + i];
+        nval = 2;
+        pos  = (32*sgitg + tiisg)*nval;
+        for (int i = 0; i < nval; ++i) shared_signs[pos+i] = ksigns_iq2xs[pos+i];
+        threadgroup_barrier(mem_flags::mem_threadgroup);
+    }
+
+#if QK_K == 256
+    const int ix = tiisg;
+
+    device const float * y4 = y + 32 * ix;
+
+    for (int ib32 = ix; ib32 < nb32; ib32 += 32) {
+
+        for (int i = 0; i < 32; ++i) {
+            yl[i] = y4[i];
+        }
+
+        const int ibl = ib32 / (QK_K / 32);
+        const int ib  = ib32 % (QK_K / 32);
+
+        device const block_iq2_xs * xr = x + ibl;
+        device const uint16_t * q2 = xr->qs + 4 * ib;
+        device const uint8_t  * sc = xr->scales + ib;
+        device const half * dh = &xr->d;
+
+        for (int row = 0; row < N_DST; row++) {
+
+            const float db = dh[0];
+            const uint8_t ls1 = sc[0] & 0xf;
+            const uint8_t ls2 = sc[0] >>  4;
+            const float d1 = db * (0.5f + ls1);
+            const float d2 = db * (0.5f + ls2);
+
+            float sum1 = 0, sum2 = 0;
+            for (int l = 0; l < 2; ++l) {
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511));
+                const uint8_t signs = shared_signs[(q2[l] >> 9)];
+                for (int j = 0; j < 8; ++j) {
+                    sum1 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+                }
+            }
+            for (int l = 2; l < 4; ++l) {
+                const threadgroup uint8_t * grid = (const threadgroup uint8_t *)(values + (q2[l] & 511));
+                const uint8_t signs = shared_signs[(q2[l] >> 9)];
+                for (int j = 0; j < 8; ++j) {
+                    sum2 += yl[8*l + j] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+                }
+            }
+            sumf[row] += d1 * sum1 + d2 * sum2;
+
+            dh += nb*sizeof(block_iq2_xs)/2;
+            q2 += nb*sizeof(block_iq2_xs)/2;
+            sc += nb*sizeof(block_iq2_xs);
+        }
+
+        y4 += 32 * 32;
+    }
+#else
+    // TODO
+#endif
+
+    for (int row = 0; row < N_DST; ++row) {
+        all_sum = simd_sum(sumf[row]);
+        if (tiisg == 0) {
+            dst[r1*ne0 + im*ne0*ne1 + first_row + row] = all_sum * 0.25f;
+        }
+    }
+}
+
+[[host_name("kernel_mul_mv_iq2_xs_f32")]]
+kernel void kernel_mul_mv_iq2_xs_f32(
+        device const  void * src0,
+        device const float * src1,
+        device       float * dst,
+        constant   int64_t & ne00,
+        constant   int64_t & ne01,
+        constant   int64_t & ne02,
+        constant  uint64_t & nb00,
+        constant  uint64_t & nb01,
+        constant  uint64_t & nb02,
+        constant   int64_t & ne10,
+        constant   int64_t & ne11,
+        constant   int64_t & ne12,
+        constant  uint64_t & nb10,
+        constant  uint64_t & nb11,
+        constant  uint64_t & nb12,
+        constant   int64_t & ne0,
+        constant   int64_t & ne1,
+        constant   uint    & r2,
+        constant   uint    & r3,
+        threadgroup int8_t * shared_values [[threadgroup(0)]],
+        uint3 tgpig[[threadgroup_position_in_grid]],
+        uint  tiisg[[thread_index_in_simdgroup]],
+        uint  sgitg[[simdgroup_index_in_threadgroup]]) {
+
+    kernel_mul_mv_iq2_xs_f32_impl(src0, src1, dst, ne00, ne01, ne02, ne10, ne12, ne0, ne1, r2, r3, shared_values, tgpig, tiisg, sgitg);
+}
+
 //============================= templates and their specializations =============================
 
 // NOTE: this is not dequantizing - we are simply fitting the template
@@ -3620,8 +4122,8 @@ void dequantize_q3_K(device const block_q3_K *xb, short il, thread type4x4 & reg
     uint16_t scale_2 = scales[il%8], scale_1 = scales[8 + il%4];
     int16_t  dl_int = (il/4)&1 ? (scale_2&kmask2) | ((scale_1&kmask1) << 2)
                                : (scale_2&kmask2) | ((scale_1&kmask1) << 4);
-    half dl = il<8 ? d_all * (dl_int - 32.h) : d_all * (dl_int / 16.h - 32.h);
-    const half ml = 4.h * dl;
+    float dl = il<8 ? d_all * (dl_int - 32.f) : d_all * (dl_int / 16.f - 32.f);
+    const float ml = 4.f * dl;
 
     il = (il/2) & 3;
     const half    coef = il>1 ? (il>2 ? 1/64.h : 1/16.h) : (il>0 ? 1/4.h : 1.h);
@@ -3688,7 +4190,7 @@ void dequantize_q5_K(device const block_q5_K *xb, short il, thread type4x4 & reg
     uint8_t ul = 1 << (il/2);
     il = il & 3;
     const uchar2 sc = get_scale_min_k4_just2(is, il/2, xb->scales);
-    const float d = il < 2 ? xb->d : xb->d / 16.h;
+    const float d = il < 2 ? xb->d : xb->d / 16.f;
     const float min = xb->dmin;
     const float dl = d * sc[0];
     const float ml = min * sc[1];
@@ -3721,17 +4223,17 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg
 #if QK_K == 256
     ql = ql + 64*(il/8) + 32*((il/2)&1) + 16*(il&1);
     qh = qh + 32*(il/8) + 16*(il&1);
-    half sc = scales[(il%2) + 2 * ((il/2))];
+    float sc = scales[(il%2) + 2 * ((il/2))];
     il = (il/2) & 3;
 #else
     ql = ql + 16 * (il&1);
-    half sc = scales[il];
+    float sc = scales[il];
 #endif
     const uint16_t  kmask1 = il>1 ? (il>2 ? 192 : 48) : (il>0 ? 12 : 3);
     const uint16_t  kmask2 = il>1 ? 0xF0              : 0x0F;
-    const half        coef = il>1 ? 1.f/16.h          : 1.h;
-    const half ml = d_all * sc * 32.h;
-    const half dl = d_all * sc * coef;
+    const float       coef = il>1 ? 1.f/16.f          : 1.f;
+    const float ml = d_all * sc * 32.f;
+    const float dl = d_all * sc * coef;
     for (int i = 0; i < 16; ++i) {
         const half q = il&1 ? ((ql[i] & kmask2) | ((qh[i] & kmask1) << 2))
                             : ((ql[i] & kmask2) | ((qh[i] & kmask1) << 4));
@@ -3739,6 +4241,52 @@ void dequantize_q6_K(device const block_q6_K *xb, short il, thread type4x4 & reg
     }
 }
 
+template <typename type4x4>
+void dequantize_iq2_xxs(device const block_iq2_xxs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    // each block of 32 needs 2 uint32_t's for the quants & scale, so 4 uint16_t's.
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const uint32_t aux32_g = q2[0] | (q2[1] << 16);
+    const uint32_t aux32_s = q2[2] | (q2[3] << 16);
+    thread const uint8_t * aux8 = (thread const uint8_t *)&aux32_g;
+    const float dl = d * (0.5f + (aux32_s >> 28)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+0]);
+    uint8_t signs = ksigns_iq2xs[(aux32_s >> 14*il) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xxs_grid + aux8[2*il+1]);
+    signs = ksigns_iq2xs[(aux32_s >> (14*il+7)) & 127];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
+template <typename type4x4>
+void dequantize_iq2_xs(device const block_iq2_xs * xb, short il, thread type4x4 & reg) {
+    // il is 0...15 for QK_K = 256 => index of block of 32 is il/2
+    const float d = xb->d;
+    const int ib32 = il/2;
+    il = il%2;
+    // il = 0 or 1. il = 0 processes the first 16 quants in a block of 32, il = 1 the second 16
+    device const uint16_t * q2 = xb->qs + 4*ib32;
+    const float dl = d * (0.5f + ((xb->scales[ib32] >> 4*il) & 0xf)) * 0.25f;
+    constant uint8_t * grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+0] & 511));
+    uint8_t signs = ksigns_iq2xs[q2[2*il+0] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+    grid = (constant uint8_t *)(iq2xs_grid + (q2[2*il+1] & 511));
+    signs = ksigns_iq2xs[q2[2*il+1] >> 9];
+    for (int i = 0; i < 8; ++i) {
+        reg[2+i/4][i%4] = dl * grid[i] * (signs & kmask_iq2xs[i] ? -1.f : 1.f);
+    }
+}
+
 template<typename block_q, short nl, void (*dequantize_func)(device const block_q *, short, thread float4x4 &)>
 kernel void kernel_get_rows(
         device const  void * src0,
@@ -4278,6 +4826,8 @@ template [[host_name("kernel_get_rows_q3_K")]] kernel get_rows_t kernel_get_rows
 template [[host_name("kernel_get_rows_q4_K")]] kernel get_rows_t kernel_get_rows<block_q4_K, QK_NL, dequantize_q4_K>;
 template [[host_name("kernel_get_rows_q5_K")]] kernel get_rows_t kernel_get_rows<block_q5_K, QK_NL, dequantize_q5_K>;
 template [[host_name("kernel_get_rows_q6_K")]] kernel get_rows_t kernel_get_rows<block_q6_K, QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_get_rows_iq2_xxs")]] kernel get_rows_t kernel_get_rows<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_get_rows_iq2_xs")]]  kernel get_rows_t kernel_get_rows<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 
 //
 // matrix-matrix multiplication
@@ -4314,6 +4864,8 @@ template [[host_name("kernel_mul_mm_q3_K_f32")]] kernel mat_mm_t kernel_mul_mm<b
 template [[host_name("kernel_mul_mm_q4_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q4_K, QK_NL, dequantize_q4_K>;
 template [[host_name("kernel_mul_mm_q5_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q5_K, QK_NL, dequantize_q5_K>;
 template [[host_name("kernel_mul_mm_q6_K_f32")]] kernel mat_mm_t kernel_mul_mm<block_q6_K, QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_mul_mm_iq2_xxs_f32")]] kernel mat_mm_t kernel_mul_mm<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_mul_mm_iq2_xs_f32")]]  kernel mat_mm_t kernel_mul_mm<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 
 //
 // indirect matrix-matrix multiplication
@@ -4362,6 +4914,8 @@ template [[host_name("kernel_mul_mm_id_q3_K_f32")]] kernel mat_mm_id_t kernel_mu
 template [[host_name("kernel_mul_mm_id_q4_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q4_K, QK_NL, dequantize_q4_K>;
 template [[host_name("kernel_mul_mm_id_q5_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q5_K, QK_NL, dequantize_q5_K>;
 template [[host_name("kernel_mul_mm_id_q6_K_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_q6_K, QK_NL, dequantize_q6_K>;
+template [[host_name("kernel_mul_mm_id_iq2_xxs_f32")]] kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xxs, QK_NL, dequantize_iq2_xxs>;
+template [[host_name("kernel_mul_mm_id_iq2_xs_f32")]]  kernel mat_mm_id_t kernel_mul_mm_id<block_iq2_xs,  QK_NL, dequantize_iq2_xs>;
 
 //
 // matrix-vector multiplication
@@ -5134,3 +5688,133 @@ kernel void kernel_mul_mv_id_q6_K_f32(
         tiisg,
         sgitg);
 }
+
+[[host_name("kernel_mul_mv_id_iq2_xxs_f32")]]
+kernel void kernel_mul_mv_id_iq2_xxs_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         float * dst,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        threadgroup int8_t   * shared_values [[threadgroup(0)]],
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_iq2_xxs_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        dst + bid*ne0,
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        shared_values,
+        tgpig,
+        tiisg,
+        sgitg);
+}
+
+[[host_name("kernel_mul_mv_id_iq2_xs_f32")]]
+kernel void kernel_mul_mv_id_iq2_xs_f32(
+        device const    char * ids,
+        device const    char * src1,
+        device         float * dst,
+        constant    uint64_t & nbi1,
+        constant     int64_t & ne00,
+        constant     int64_t & ne01,
+        constant     int64_t & ne02,
+        constant    uint64_t & nb00,
+        constant    uint64_t & nb01,
+        constant    uint64_t & nb02,
+        constant     int64_t & ne10,
+        constant     int64_t & ne11,
+        constant     int64_t & ne12,
+        constant     int64_t & ne13,
+        constant    uint64_t & nb10,
+        constant    uint64_t & nb11,
+        constant    uint64_t & nb12,
+        constant     int64_t & ne0,
+        constant     int64_t & ne1,
+        constant    uint64_t & nb1,
+        constant        uint & r2,
+        constant        uint & r3,
+        constant         int & idx,
+        device const    char * src00,
+        device const    char * src01,
+        device const    char * src02,
+        device const    char * src03,
+        device const    char * src04,
+        device const    char * src05,
+        device const    char * src06,
+        device const    char * src07,
+        threadgroup int8_t   * shared_values [[threadgroup(0)]],
+        uint3                  tgpig[[threadgroup_position_in_grid]],
+        uint                   tiitg[[thread_index_in_threadgroup]],
+        uint                   tiisg[[thread_index_in_simdgroup]],
+        uint                   sgitg[[simdgroup_index_in_threadgroup]]) {
+    device const char * src0[8] = {src00, src01, src02, src03, src04, src05, src06, src07};
+
+    const int64_t bid = tgpig.z/(ne12*ne13);
+
+    tgpig.z = tgpig.z%(ne12*ne13);
+
+    const int32_t id = ((device int32_t *) (ids + bid*nbi1))[idx];
+
+    kernel_mul_mv_iq2_xs_f32_impl(
+        src0[id],
+        (device const float *) (src1 + bid*nb11),
+        dst + bid*ne0,
+        ne00,
+        ne01,
+        ne02,
+        ne10,
+        ne12,
+        ne0,
+        ne1,
+        r2,
+        r3,
+        shared_values,
+        tgpig,
+        tiisg,
+        sgitg);
+}

ggml-quants.c

@@ -2340,6 +2340,322 @@ size_t ggml_quantize_q6_K(const float * src, void * dst, int n, int k, int64_t *
     return (n/QK_K*sizeof(block_q6_K));
 }
 
+// ====================== "True" 2-bit (de)-quantization
+
+static const  uint64_t iq2xxs_grid[256] = {
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x08080808082b0808,
+    0x08080808082b082b, 0x08080808082b2b08, 0x08080808082b2b2b, 0x0808080819080819,
+    0x0808080819081908, 0x0808080819190808, 0x0808080819192b08, 0x08080808192b0819,
+    0x08080808192b1908, 0x080808082b080808, 0x080808082b08082b, 0x080808082b082b2b,
+    0x080808082b2b082b, 0x0808081908080819, 0x0808081908081908, 0x0808081908190808,
+    0x0808081908191919, 0x0808081919080808, 0x080808192b081908, 0x080808192b192b08,
+    0x0808082b08080808, 0x0808082b0808082b, 0x0808082b082b082b, 0x0808082b2b08082b,
+    0x0808190808080819, 0x0808190808081908, 0x0808190808190808, 0x08081908082b0819,
+    0x08081908082b1908, 0x0808190819080808, 0x080819081908082b, 0x0808190819082b08,
+    0x08081908192b0808, 0x080819082b080819, 0x080819082b081908, 0x080819082b190808,
+    0x080819082b2b1908, 0x0808191908080808, 0x080819190808082b, 0x0808191908082b08,
+    0x08081919082b0808, 0x080819191908192b, 0x08081919192b2b19, 0x080819192b080808,
+    0x080819192b190819, 0x0808192b08082b19, 0x0808192b08190808, 0x0808192b19080808,
+    0x0808192b2b081908, 0x0808192b2b2b1908, 0x08082b0808080808, 0x08082b0808081919,
+    0x08082b0808082b08, 0x08082b0808191908, 0x08082b08082b2b08, 0x08082b0819080819,
+    0x08082b0819081908, 0x08082b0819190808, 0x08082b081919082b, 0x08082b082b082b08,
+    0x08082b1908081908, 0x08082b1919080808, 0x08082b2b0808082b, 0x08082b2b08191908,
+    0x0819080808080819, 0x0819080808081908, 0x0819080808190808, 0x08190808082b0819,
+    0x0819080819080808, 0x08190808192b0808, 0x081908082b081908, 0x081908082b190808,
+    0x081908082b191919, 0x0819081908080808, 0x0819081908082b08, 0x08190819082b0808,
+    0x0819081919190808, 0x0819081919192b2b, 0x081908192b080808, 0x0819082b082b1908,
+    0x0819082b19081919, 0x0819190808080808, 0x0819190808082b08, 0x08191908082b0808,
+    0x08191908082b1919, 0x0819190819082b19, 0x081919082b080808, 0x0819191908192b08,
+    0x08191919192b082b, 0x0819192b08080808, 0x0819192b0819192b, 0x08192b0808080819,
+    0x08192b0808081908, 0x08192b0808190808, 0x08192b0819080808, 0x08192b082b080819,
+    0x08192b1908080808, 0x08192b1908081919, 0x08192b192b2b0808, 0x08192b2b19190819,
+    0x082b080808080808, 0x082b08080808082b, 0x082b080808082b2b, 0x082b080819081908,
+    0x082b0808192b0819, 0x082b08082b080808, 0x082b08082b08082b, 0x082b0819082b2b19,
+    0x082b081919082b08, 0x082b082b08080808, 0x082b082b0808082b, 0x082b190808080819,
+    0x082b190808081908, 0x082b190808190808, 0x082b190819080808, 0x082b19081919192b,
+    0x082b191908080808, 0x082b191919080819, 0x082b1919192b1908, 0x082b192b2b190808,
+    0x082b2b0808082b08, 0x082b2b08082b0808, 0x082b2b082b191908, 0x082b2b2b19081908,
+    0x1908080808080819, 0x1908080808081908, 0x1908080808190808, 0x1908080808192b08,
+    0x19080808082b0819, 0x19080808082b1908, 0x1908080819080808, 0x1908080819082b08,
+    0x190808081919192b, 0x19080808192b0808, 0x190808082b080819, 0x190808082b081908,
+    0x190808082b190808, 0x1908081908080808, 0x19080819082b0808, 0x19080819192b0819,
+    0x190808192b080808, 0x190808192b081919, 0x1908082b08080819, 0x1908082b08190808,
+    0x1908082b19082b08, 0x1908082b1919192b, 0x1908082b192b2b08, 0x1908190808080808,
+    0x1908190808082b08, 0x19081908082b0808, 0x190819082b080808, 0x190819082b192b19,
+    0x190819190819082b, 0x19081919082b1908, 0x1908192b08080808, 0x19082b0808080819,
+    0x19082b0808081908, 0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919,
+    0x19082b1908080808, 0x19082b1919192b08, 0x19082b19192b0819, 0x19082b192b08082b,
+    0x19082b2b19081919, 0x19082b2b2b190808, 0x1919080808080808, 0x1919080808082b08,
+    0x1919080808190819, 0x1919080808192b19, 0x19190808082b0808, 0x191908082b080808,
+    0x191908082b082b08, 0x1919081908081908, 0x191908191908082b, 0x191908192b2b1908,
+    0x1919082b2b190819, 0x191919082b190808, 0x191919082b19082b, 0x1919191908082b2b,
+    0x1919192b08080819, 0x1919192b19191908, 0x19192b0808080808, 0x19192b0808190819,
+    0x19192b0808192b19, 0x19192b08192b1908, 0x19192b1919080808, 0x19192b2b08082b08,
+    0x192b080808081908, 0x192b080808190808, 0x192b080819080808, 0x192b0808192b2b08,
+    0x192b081908080808, 0x192b081919191919, 0x192b082b08192b08, 0x192b082b192b0808,
+    0x192b190808080808, 0x192b190808081919, 0x192b191908190808, 0x192b19190819082b,
+    0x192b19192b081908, 0x192b2b081908082b, 0x2b08080808080808, 0x2b0808080808082b,
+    0x2b08080808082b2b, 0x2b08080819080819, 0x2b0808082b08082b, 0x2b08081908081908,
+    0x2b08081908192b08, 0x2b08081919080808, 0x2b08082b08190819, 0x2b08190808080819,
+    0x2b08190808081908, 0x2b08190808190808, 0x2b08190808191919, 0x2b08190819080808,
+    0x2b081908192b0808, 0x2b08191908080808, 0x2b0819191908192b, 0x2b0819192b191908,
+    0x2b08192b08082b19, 0x2b08192b19080808, 0x2b08192b192b0808, 0x2b082b080808082b,
+    0x2b082b1908081908, 0x2b082b2b08190819, 0x2b19080808081908, 0x2b19080808190808,
+    0x2b190808082b1908, 0x2b19080819080808, 0x2b1908082b2b0819, 0x2b1908190819192b,
+    0x2b1908192b080808, 0x2b19082b19081919, 0x2b19190808080808, 0x2b191908082b082b,
+    0x2b19190819081908, 0x2b19191919190819, 0x2b192b082b080819, 0x2b192b19082b0808,
+    0x2b2b08080808082b, 0x2b2b080819190808, 0x2b2b08082b081919, 0x2b2b081908082b19,
+    0x2b2b082b08080808, 0x2b2b190808192b08, 0x2b2b2b0819190808, 0x2b2b2b1908081908,
+};
+
+static const uint64_t iq2xs_grid[512] = {
+    0x0808080808080808, 0x080808080808082b, 0x0808080808081919, 0x0808080808082b08,
+    0x0808080808082b2b, 0x0808080808190819, 0x0808080808191908, 0x080808080819192b,
+    0x0808080808192b19, 0x08080808082b0808, 0x08080808082b082b, 0x08080808082b1919,
+    0x08080808082b2b08, 0x0808080819080819, 0x0808080819081908, 0x080808081908192b,
+    0x0808080819082b19, 0x0808080819190808, 0x080808081919082b, 0x0808080819191919,
+    0x0808080819192b08, 0x08080808192b0819, 0x08080808192b1908, 0x080808082b080808,
+    0x080808082b08082b, 0x080808082b081919, 0x080808082b082b08, 0x080808082b190819,
+    0x080808082b191908, 0x080808082b192b19, 0x080808082b2b0808, 0x0808081908080819,
+    0x0808081908081908, 0x080808190808192b, 0x0808081908082b19, 0x0808081908190808,
+    0x080808190819082b, 0x0808081908191919, 0x0808081908192b08, 0x0808081908192b2b,
+    0x08080819082b0819, 0x08080819082b1908, 0x0808081919080808, 0x080808191908082b,
+    0x0808081919081919, 0x0808081919082b08, 0x0808081919190819, 0x0808081919191908,
+    0x08080819192b0808, 0x08080819192b2b08, 0x080808192b080819, 0x080808192b081908,
+    0x080808192b190808, 0x0808082b08080808, 0x0808082b0808082b, 0x0808082b08081919,
+    0x0808082b08082b08, 0x0808082b08190819, 0x0808082b08191908, 0x0808082b082b0808,
+    0x0808082b19080819, 0x0808082b19081908, 0x0808082b19190808, 0x0808082b19191919,
+    0x0808082b2b080808, 0x0808082b2b082b2b, 0x0808190808080819, 0x0808190808081908,
+    0x080819080808192b, 0x0808190808082b19, 0x0808190808190808, 0x080819080819082b,
+    0x0808190808191919, 0x0808190808192b08, 0x08081908082b0819, 0x08081908082b1908,
+    0x0808190819080808, 0x080819081908082b, 0x0808190819081919, 0x0808190819082b08,
+    0x0808190819190819, 0x0808190819191908, 0x080819081919192b, 0x08081908192b0808,
+    0x080819082b080819, 0x080819082b081908, 0x080819082b190808, 0x0808191908080808,
+    0x080819190808082b, 0x0808191908081919, 0x0808191908082b08, 0x0808191908190819,
+    0x0808191908191908, 0x08081919082b0808, 0x0808191919080819, 0x0808191919081908,
+    0x0808191919190808, 0x08081919192b0819, 0x080819192b080808, 0x0808192b08080819,
+    0x0808192b08081908, 0x0808192b08190808, 0x0808192b082b192b, 0x0808192b19080808,
+    0x0808192b1908082b, 0x0808192b2b081908, 0x08082b0808080808, 0x08082b080808082b,
+    0x08082b0808081919, 0x08082b0808082b08, 0x08082b0808082b2b, 0x08082b0808190819,
+    0x08082b0808191908, 0x08082b08082b0808, 0x08082b08082b1919, 0x08082b0819080819,
+    0x08082b0819081908, 0x08082b0819190808, 0x08082b0819192b08, 0x08082b082b080808,
+    0x08082b082b2b0808, 0x08082b082b2b2b2b, 0x08082b1908080819, 0x08082b1908081908,
+    0x08082b1908190808, 0x08082b1919080808, 0x08082b192b080819, 0x08082b192b082b19,
+    0x08082b2b08080808, 0x08082b2b082b0808, 0x08082b2b082b2b08, 0x08082b2b2b19192b,
+    0x08082b2b2b2b0808, 0x0819080808080819, 0x0819080808081908, 0x081908080808192b,
+    0x0819080808082b19, 0x0819080808190808, 0x081908080819082b, 0x0819080808191919,
+    0x0819080808192b08, 0x08190808082b0819, 0x08190808082b1908, 0x0819080819080808,
+    0x081908081908082b, 0x0819080819081919, 0x0819080819082b08, 0x0819080819190819,
+    0x0819080819191908, 0x08190808192b0808, 0x08190808192b2b2b, 0x081908082b080819,
+    0x081908082b081908, 0x081908082b190808, 0x0819081908080808, 0x081908190808082b,
+    0x0819081908081919, 0x0819081908082b08, 0x0819081908190819, 0x0819081908191908,
+    0x08190819082b0808, 0x0819081919080819, 0x0819081919081908, 0x0819081919190808,
+    0x081908192b080808, 0x081908192b191908, 0x081908192b19192b, 0x0819082b08080819,
+    0x0819082b08081908, 0x0819082b0808192b, 0x0819082b08190808, 0x0819082b19080808,
+    0x0819082b192b0808, 0x0819190808080808, 0x081919080808082b, 0x0819190808081919,
+    0x0819190808082b08, 0x0819190808190819, 0x0819190808191908, 0x08191908082b0808,
+    0x0819190819080819, 0x0819190819081908, 0x0819190819082b19, 0x0819190819190808,
+    0x08191908192b1908, 0x081919082b080808, 0x0819191908080819, 0x0819191908081908,
+    0x0819191908190808, 0x0819191919080808, 0x0819192b08080808, 0x0819192b08191908,
+    0x0819192b19082b19, 0x08192b0808080819, 0x08192b0808081908, 0x08192b0808190808,
+    0x08192b080819082b, 0x08192b0819080808, 0x08192b0819191908, 0x08192b082b08192b,
+    0x08192b1908080808, 0x08192b1908081919, 0x08192b19192b192b, 0x08192b2b19190819,
+    0x08192b2b2b2b2b19, 0x082b080808080808, 0x082b08080808082b, 0x082b080808081919,
+    0x082b080808082b08, 0x082b080808082b2b, 0x082b080808190819, 0x082b080808191908,
+    0x082b0808082b0808, 0x082b080819080819, 0x082b080819081908, 0x082b080819190808,
+    0x082b08082b080808, 0x082b08082b2b0808, 0x082b081908080819, 0x082b081908081908,
+    0x082b081908190808, 0x082b081919080808, 0x082b081919082b08, 0x082b0819192b1919,
+    0x082b082b08080808, 0x082b082b082b082b, 0x082b082b2b080808, 0x082b082b2b2b2b08,
+    0x082b190808080819, 0x082b190808081908, 0x082b190808190808, 0x082b1908082b2b19,
+    0x082b190819080808, 0x082b191908080808, 0x082b191919080819, 0x082b19191919082b,
+    0x082b19192b192b19, 0x082b192b08080819, 0x082b192b08192b2b, 0x082b192b2b2b192b,
+    0x082b2b0808080808, 0x082b2b0808082b08, 0x082b2b0808082b2b, 0x082b2b08082b0808,
+    0x082b2b0819191919, 0x082b2b082b082b08, 0x082b2b082b2b082b, 0x082b2b19192b2b08,
+    0x082b2b192b190808, 0x082b2b2b08082b08, 0x082b2b2b082b0808, 0x082b2b2b2b08082b,
+    0x082b2b2b2b082b08, 0x082b2b2b2b082b2b, 0x1908080808080819, 0x1908080808081908,
+    0x190808080808192b, 0x1908080808082b19, 0x1908080808190808, 0x190808080819082b,
+    0x1908080808191919, 0x1908080808192b08, 0x19080808082b0819, 0x19080808082b1908,
+    0x1908080819080808, 0x190808081908082b, 0x1908080819081919, 0x1908080819082b08,
+    0x1908080819082b2b, 0x1908080819190819, 0x1908080819191908, 0x19080808192b0808,
+    0x19080808192b1919, 0x190808082b080819, 0x190808082b081908, 0x190808082b190808,
+    0x1908081908080808, 0x190808190808082b, 0x1908081908081919, 0x1908081908082b08,
+    0x1908081908190819, 0x1908081908191908, 0x19080819082b0808, 0x1908081919080819,
+    0x1908081919081908, 0x1908081919190808, 0x190808192b080808, 0x190808192b081919,
+    0x190808192b2b082b, 0x1908082b08080819, 0x1908082b08081908, 0x1908082b08190808,
+    0x1908082b0819082b, 0x1908082b082b2b19, 0x1908082b19080808, 0x1908190808080808,
+    0x190819080808082b, 0x1908190808081919, 0x1908190808082b08, 0x1908190808190819,
+    0x1908190808191908, 0x1908190808192b19, 0x19081908082b0808, 0x1908190819080819,
+    0x1908190819081908, 0x1908190819190808, 0x190819082b080808, 0x190819082b191908,
+    0x1908191908080819, 0x1908191908081908, 0x1908191908190808, 0x19081919082b1908,
+    0x1908191919080808, 0x190819192b192b2b, 0x1908192b08080808, 0x1908192b08082b2b,
+    0x1908192b19081908, 0x1908192b19190808, 0x19082b0808080819, 0x19082b0808081908,
+    0x19082b0808190808, 0x19082b0819080808, 0x19082b0819081919, 0x19082b0819191908,
+    0x19082b08192b082b, 0x19082b1908080808, 0x19082b1908190819, 0x19082b1919081908,
+    0x19082b1919190808, 0x19082b19192b2b19, 0x19082b2b08081908, 0x1919080808080808,
+    0x191908080808082b, 0x1919080808081919, 0x1919080808082b08, 0x1919080808190819,
+    0x1919080808191908, 0x19190808082b0808, 0x19190808082b2b08, 0x1919080819080819,
+    0x1919080819081908, 0x1919080819190808, 0x191908082b080808, 0x1919081908080819,
+    0x1919081908081908, 0x1919081908190808, 0x1919081908191919, 0x1919081919080808,
+    0x191908191908082b, 0x1919082b08080808, 0x1919082b19081908, 0x1919082b2b2b2b2b,
+    0x1919190808080819, 0x1919190808081908, 0x1919190808190808, 0x19191908082b0819,
+    0x1919190819080808, 0x19191908192b0808, 0x191919082b080819, 0x191919082b2b0819,
+    0x1919191908080808, 0x1919191908082b08, 0x191919192b080808, 0x191919192b082b08,
+    0x1919192b082b0819, 0x1919192b192b2b08, 0x1919192b2b2b0819, 0x19192b0808080808,
+    0x19192b0808191908, 0x19192b0819080819, 0x19192b0819190808, 0x19192b082b192b19,
+    0x19192b1908192b2b, 0x19192b1919080808, 0x19192b191908082b, 0x19192b2b2b081919,
+    0x192b080808080819, 0x192b080808081908, 0x192b080808190808, 0x192b080819080808,
+    0x192b080819191908, 0x192b0808192b082b, 0x192b08082b08192b, 0x192b08082b2b2b19,
+    0x192b081908080808, 0x192b082b082b1908, 0x192b082b19082b2b, 0x192b082b2b19082b,
+    0x192b190808080808, 0x192b19080819192b, 0x192b191908190808, 0x192b191919080808,
+    0x192b191919081919, 0x192b19192b2b1908, 0x192b2b0808080819, 0x192b2b08192b2b2b,
+    0x192b2b19082b1919, 0x192b2b2b0808192b, 0x192b2b2b19191908, 0x192b2b2b192b082b,
+    0x2b08080808080808, 0x2b0808080808082b, 0x2b08080808081919, 0x2b08080808082b08,
+    0x2b08080808190819, 0x2b08080808191908, 0x2b080808082b0808, 0x2b080808082b2b2b,
+    0x2b08080819080819, 0x2b08080819081908, 0x2b08080819190808, 0x2b0808082b080808,
+    0x2b0808082b08082b, 0x2b0808082b2b2b08, 0x2b0808082b2b2b2b, 0x2b08081908080819,
+    0x2b08081908081908, 0x2b0808190808192b, 0x2b08081908190808, 0x2b08081919080808,
+    0x2b08081919190819, 0x2b08081919192b19, 0x2b08082b08080808, 0x2b08082b082b0808,
+    0x2b08082b2b080808, 0x2b08082b2b08082b, 0x2b08082b2b2b0808, 0x2b08082b2b2b2b08,
+    0x2b08190808080819, 0x2b08190808081908, 0x2b08190808190808, 0x2b0819080819082b,
+    0x2b08190808191919, 0x2b08190819080808, 0x2b081908192b0808, 0x2b0819082b082b19,
+    0x2b08191908080808, 0x2b08191919081908, 0x2b0819192b2b1919, 0x2b08192b08192b08,
+    0x2b08192b192b2b2b, 0x2b082b0808080808, 0x2b082b0808082b08, 0x2b082b08082b1919,
+    0x2b082b0819192b2b, 0x2b082b082b080808, 0x2b082b082b08082b, 0x2b082b082b2b2b08,
+    0x2b082b190808192b, 0x2b082b2b082b082b, 0x2b082b2b2b080808, 0x2b082b2b2b082b08,
+    0x2b082b2b2b19192b, 0x2b082b2b2b2b2b08, 0x2b19080808080819, 0x2b19080808081908,
+    0x2b19080808190808, 0x2b19080819080808, 0x2b1908081919192b, 0x2b1908082b081908,
+    0x2b19081908080808, 0x2b190819082b082b, 0x2b190819192b1908, 0x2b19082b1919192b,
+    0x2b19082b2b082b19, 0x2b19190808080808, 0x2b19190808081919, 0x2b19190819081908,
+    0x2b19190819190808, 0x2b19190819192b08, 0x2b191919082b2b19, 0x2b1919192b190808,
+    0x2b1919192b19082b, 0x2b19192b19080819, 0x2b192b0819190819, 0x2b192b082b2b192b,
+    0x2b192b1919082b19, 0x2b192b2b08191919, 0x2b192b2b192b0808, 0x2b2b080808080808,
+    0x2b2b08080808082b, 0x2b2b080808082b08, 0x2b2b080808082b2b, 0x2b2b0808082b0808,
+    0x2b2b0808082b2b2b, 0x2b2b08082b2b0808, 0x2b2b081919190819, 0x2b2b081919192b19,
+    0x2b2b08192b2b192b, 0x2b2b082b08080808, 0x2b2b082b0808082b, 0x2b2b082b08082b08,
+    0x2b2b082b082b2b2b, 0x2b2b082b2b080808, 0x2b2b082b2b2b0808, 0x2b2b190819080808,
+    0x2b2b19082b191919, 0x2b2b192b192b1919, 0x2b2b192b2b192b08, 0x2b2b2b0808082b2b,
+    0x2b2b2b08082b0808, 0x2b2b2b08082b082b, 0x2b2b2b08082b2b08, 0x2b2b2b082b2b0808,
+    0x2b2b2b082b2b2b08, 0x2b2b2b1908081908, 0x2b2b2b192b081908, 0x2b2b2b192b08192b,
+    0x2b2b2b2b082b2b08, 0x2b2b2b2b082b2b2b, 0x2b2b2b2b2b190819, 0x2b2b2b2b2b2b2b2b,
+};
+
+static const uint8_t ksigns_iq2xs[128] = {
+      0, 129, 130,   3, 132,   5,   6, 135, 136,   9,  10, 139,  12, 141, 142,  15,
+    144,  17,  18, 147,  20, 149, 150,  23,  24, 153, 154,  27, 156,  29,  30, 159,
+    160,  33,  34, 163,  36, 165, 166,  39,  40, 169, 170,  43, 172,  45,  46, 175,
+     48, 177, 178,  51, 180,  53,  54, 183, 184,  57,  58, 187,  60, 189, 190,  63,
+    192,  65,  66, 195,  68, 197, 198,  71,  72, 201, 202,  75, 204,  77,  78, 207,
+     80, 209, 210,  83, 212,  85,  86, 215, 216,  89,  90, 219,  92, 221, 222,  95,
+     96, 225, 226,  99, 228, 101, 102, 231, 232, 105, 106, 235, 108, 237, 238, 111,
+    240, 113, 114, 243, 116, 245, 246, 119, 120, 249, 250, 123, 252, 125, 126, 255,
+};
+
+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;
+
+    uint32_t aux32[2];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(aux32, x[i].qs + 4*ib32, 2*sizeof(uint32_t));
+            const float db = d * (0.5f + (aux32[1] >> 28)) * 0.25f;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+                for (int j = 0; j < 8; ++j) {
+                    y[j] = db * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+                }
+                y += 8;
+            }
+        }
+    }
+}
+
+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;
+
+    float db[2];
+
+    for (int i = 0; i < nb; i++) {
+
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            db[0] = d * (0.5f + (x[i].scales[ib32] & 0xf)) * 0.25f;
+            db[1] = d * (0.5f + (x[i].scales[ib32] >>  4)) * 0.25f;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (x[i].qs[4*ib32 + l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[x[i].qs[4*ib32 + l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    y[j] = db[l/2] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+                }
+                y += 8;
+            }
+        }
+    }
+}
+
+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) {
@@ -2362,7 +2678,9 @@ void quantize_row_q8_K_reference(const float * restrict x, block_q8_K * restrict
             x += QK_K;
             continue;
         }
-        const float iscale = -128.f/max;
+        //const float iscale = -128.f/max;
+        // We need this change for IQ2_XXS, else the AVX implementation becomes very awkward
+        const float iscale = -127.f/max;
         for (int j = 0; j < QK_K; ++j) {
             int v = nearest_int(iscale*x[j]);
             y[i].qs[j] = MIN(127, v);
@@ -7065,3 +7383,319 @@ void ggml_vec_dot_q6_K_q8_K(const int n, float * restrict s, const void * restri
 }
 
 #endif
+
+static const int8_t keven_signs_q2xs[1024] = {
+     1,  1,  1,  1,  1,  1,  1,  1, -1,  1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1,  1,
+     1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1,  1,  1, -1, -1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1, -1,
+     1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1,  1,  1, -1,  1, -1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1, -1,
+     1,  1, -1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1,  1,
+     1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1, -1,
+     1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1,  1,
+     1,  1,  1, -1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1,  1,
+     1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1,  1,  1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1, -1,
+     1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1,  1, -1,  1,  1, -1, -1,  1,  1,  1, -1,  1, -1,
+     1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1,  1,
+     1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1,  1,
+     1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1, -1,
+     1,  1,  1,  1, -1, -1,  1,  1, -1,  1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1,  1,
+     1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1, -1,
+     1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1,  1,  1,  1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1, -1,
+     1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1,  1,
+     1,  1,  1,  1,  1,  1, -1, -1, -1,  1,  1,  1,  1,  1, -1,  1,  1, -1,  1,  1,  1,  1, -1,  1, -1, -1,  1,  1,  1,  1, -1, -1,
+     1,  1, -1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1, -1,  1, -1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1,  1,
+     1,  1,  1, -1,  1,  1, -1,  1, -1,  1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1,  1,
+     1,  1, -1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1, -1,
+     1,  1,  1,  1, -1,  1, -1,  1, -1,  1,  1,  1, -1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1,  1,
+     1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1,  1,  1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1, -1,
+     1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1,  1, -1,  1,  1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1, -1,
+     1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1,  1,
+     1,  1,  1,  1,  1, -1, -1,  1, -1,  1,  1,  1,  1, -1, -1, -1,  1, -1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1,  1,
+     1,  1, -1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1,  1, -1, -1,  1, -1, -1, -1,  1,  1, -1, -1, -1,
+     1,  1,  1, -1,  1, -1, -1, -1, -1,  1,  1, -1,  1, -1, -1,  1,  1, -1,  1, -1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1, -1,
+     1,  1, -1, -1,  1, -1, -1,  1, -1,  1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1,  1,
+     1,  1,  1,  1, -1, -1, -1, -1, -1,  1,  1,  1, -1, -1, -1,  1,  1, -1,  1,  1, -1, -1, -1,  1, -1, -1,  1,  1, -1, -1, -1, -1,
+     1,  1, -1,  1, -1, -1, -1,  1, -1,  1, -1,  1, -1, -1, -1, -1,  1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1,  1,
+     1,  1,  1, -1, -1, -1, -1,  1, -1,  1,  1, -1, -1, -1, -1, -1,  1, -1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1,  1,
+     1,  1, -1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1,  1,  1, -1, -1, -1, -1, -1, -1,  1, -1, -1, -1, -1, -1, -1, -1, -1,
+};
+
+void ggml_vec_dot_iq2_xxs_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+    assert(n % QK_K == 0);
+
+    const block_iq2_xxs * restrict x = vx;
+    const block_q8_K    * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    ggml_int8x16x4_t q2u;
+    ggml_int8x16x4_t q2s;
+    ggml_int8x16x4_t q8b;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        float sumf1 = 0, sumf2 = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            q8b = ggml_vld1q_s8_x4(q8); q8 += 64;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+            q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 0])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 1])));
+            q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 2])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 3])));
+            q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[ 8])), vld1_s8((const void *)(iq2xxs_grid + aux8[ 9])));
+            q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xxs_grid + aux8[10])), vld1_s8((const void *)(iq2xxs_grid + aux8[11])));
+            q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >>  7) & 127))));
+            q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[1] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[1] >> 21) & 127))));
+            q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[3] >>  0) & 127))), vld1_s8((const void *)(signs64 + ((aux32[3] >>  7) & 127))));
+            q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + ((aux32[3] >> 14) & 127))), vld1_s8((const void *)(signs64 + ((aux32[3] >> 21) & 127))));
+            q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]);
+            q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]);
+            q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]);
+            q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]);
+            const int32x4_t p1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]), q2u.val[1], q8b.val[1]);
+            const int32x4_t p2 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]), q2u.val[3], q8b.val[3]);
+            sumf1 += vaddvq_s32(p1) * (0.5f + (aux32[1] >> 28));
+            sumf2 += vaddvq_s32(p2) * (0.5f + (aux32[3] >> 28));
+        }
+        sumf += d*(sumf1 + sumf2);
+    }
+    *s = 0.25f * sumf;
+
+#elif defined(__AVX2__)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint32_t aux32[4];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+            const __m256i q2_1 = _mm256_set_epi64x(iq2xxs_grid[aux8[ 3]], iq2xxs_grid[aux8[ 2]], iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
+            const __m256i q2_2 = _mm256_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]], iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
+            const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+                                                   signs64[(aux32[1] >>  7) & 127], signs64[(aux32[1] >>  0) & 127]);
+            const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127],
+                                                   signs64[(aux32[3] >>  7) & 127], signs64[(aux32[3] >>  0) & 127]);
+            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
+            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+            const uint16_t ls1 = aux32[1] >> 28;
+            const uint16_t ls2 = aux32[3] >> 28;
+            const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
+            const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
+            sumi1 = _mm256_add_epi32(sumi1, p1);
+            sumi2 = _mm256_add_epi32(sumi2, p2);
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#else
+
+    uint32_t aux32[2];
+    const uint8_t * aux8 = (const uint8_t *)aux32;
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            memcpy(aux32, q2, 2*sizeof(uint32_t));
+            q2 += 4;
+            const uint32_t ls = 2*(aux32[1] >> 28) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+                const uint8_t  signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+#endif
+}
+
+void ggml_vec_dot_iq2_xs_q8_K(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+    assert(n % QK_K == 0);
+
+    const block_iq2_xs * restrict x = vx;
+    const block_q8_K   * restrict y = vy;
+
+    const int nb = n / QK_K;
+
+#if defined(__ARM_NEON)
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    int8x16x4_t q2u;
+    int8x16x4_t q2s;
+    int8x16x4_t q8b;
+
+    int32x4x4_t scales32;
+
+    float sumf = 0;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+        const uint8x8_t scales8 = vld1_u8(x[i].scales);
+        const uint8x8_t scales_l = vand_u8(scales8, vdup_n_u8(0xf));
+        const uint8x8_t scales_h = vshr_n_u8(scales8, 4);
+        uint8x16_t scales = vcombine_u8(vzip1_u8(scales_l, scales_h), vzip2_u8(scales_l, scales_h));
+        scales = vaddq_u8(vshlq_n_u8(scales, 1), vdupq_n_u8(1));
+        const uint16x8_t scales1 = vmovl_u8(vget_low_u8(scales));
+        const uint16x8_t scales2 = vmovl_u8(vget_high_u8(scales));
+        scales32.val[0] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales1)));
+        scales32.val[1] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales1)));
+        scales32.val[2] = vreinterpretq_s32_u32(vmovl_u16(vget_low_u16(scales2)));
+        scales32.val[3] = vreinterpretq_s32_u32(vmovl_u16(vget_high_u16(scales2)));
+        int32x4_t sumi = vdupq_n_s32(0);
+        for (int ib64 = 0; ib64 < QK_K/64; ++ib64) {
+            q8b = vld1q_s8_x4(q8); q8 += 64;
+            q2u.val[0] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[0] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[1] & 511))));
+            q2u.val[1] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[2] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[3] & 511))));
+            q2u.val[2] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[4] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[5] & 511))));
+            q2u.val[3] = vcombine_s8(vld1_s8((const void *)(iq2xs_grid + (q2[6] & 511))), vld1_s8((const void *)(iq2xs_grid + (q2[7] & 511))));
+            q2s.val[0] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[0] >> 9))), vld1_s8((const void *)(signs64 + (q2[1] >> 9))));
+            q2s.val[1] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[2] >> 9))), vld1_s8((const void *)(signs64 + (q2[3] >> 9))));
+            q2s.val[2] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[4] >> 9))), vld1_s8((const void *)(signs64 + (q2[5] >> 9))));
+            q2s.val[3] = vcombine_s8(vld1_s8((const void *)(signs64 + (q2[6] >> 9))), vld1_s8((const void *)(signs64 + (q2[7] >> 9))));
+            q2u.val[0] = vmulq_s8(q2u.val[0], q2s.val[0]);
+            q2u.val[1] = vmulq_s8(q2u.val[1], q2s.val[1]);
+            q2u.val[2] = vmulq_s8(q2u.val[2], q2s.val[2]);
+            q2u.val[3] = vmulq_s8(q2u.val[3], q2s.val[3]);
+            const int32x4_t p1 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[0], q8b.val[0]);
+            const int32x4_t p2 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[1], q8b.val[1]);
+            const int32x4_t p3 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[2], q8b.val[2]);
+            const int32x4_t p4 = ggml_vdotq_s32(vdupq_n_s32(0), q2u.val[3], q8b.val[3]);
+            const int32x4_t p = vpaddq_s32(vpaddq_s32(p1, p2), vpaddq_s32(p3, p4));
+            sumi = vmlaq_s32(sumi, p, scales32.val[ib64]);
+            q2 += 8;
+        }
+        sumf += d*vaddvq_s32(sumi);
+    }
+    *s = 0.125f * sumf;
+
+#elif defined(__AVX2__)
+
+    const __m128i m4 = _mm_set1_epi8(0xf);
+    const __m128i m1 = _mm_set1_epi8(1);
+    const __m128i m511 = _mm_set1_epi16(511);
+    const __m128i m127 = _mm_set1_epi16(127);
+
+    const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+    uint64_t aux64;
+
+    // somewhat hacky, but gives a significant boost in performance
+    __m128i aux_gindex, aux_sindex;
+    const uint16_t * gindex = (const uint16_t *)&aux_gindex;
+    const uint16_t * sindex = (const uint16_t *)&aux_sindex;
+
+    __m256 accumf = _mm256_setzero_ps();
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const int8_t   * restrict q8 = y[i].qs;
+
+        memcpy(&aux64, x[i].scales, 8);
+        __m128i stmp = _mm_set1_epi64x(aux64);
+        stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
+        const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
+
+        __m256i sumi1 = _mm256_setzero_si256();
+        __m256i sumi2 = _mm256_setzero_si256();
+        for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+            const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+            const __m128i q2_data = _mm_loadu_si128((const __m128i*)q2);  q2 +=  8;
+            aux_gindex = _mm_and_si128(q2_data, m511);
+            aux_sindex = _mm_and_si128(_mm_srli_epi16(q2_data, 9), m127);
+            const __m256i q2_1 = _mm256_set_epi64x(iq2xs_grid[gindex[3]], iq2xs_grid[gindex[2]], iq2xs_grid[gindex[1]], iq2xs_grid[gindex[0]]);
+            const __m256i q2_2 = _mm256_set_epi64x(iq2xs_grid[gindex[7]], iq2xs_grid[gindex[6]], iq2xs_grid[gindex[5]], iq2xs_grid[gindex[4]]);
+            const __m256i s2_1 = _mm256_set_epi64x(signs64[sindex[3]], signs64[sindex[2]], signs64[sindex[1]], signs64[sindex[0]]);
+            const __m256i s2_2 = _mm256_set_epi64x(signs64[sindex[7]], signs64[sindex[6]], signs64[sindex[5]], signs64[sindex[4]]);
+            const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
+            const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
+            const __m256i dot1  = _mm256_maddubs_epi16(q2_1, q8s_1);
+            const __m256i dot2  = _mm256_maddubs_epi16(q2_2, q8s_2);
+
+            const __m256i sc1 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0)));
+            const __m256i sc2 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1)));
+
+            sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot1, sc1));
+            sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot2, sc2));
+        }
+
+        accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+    }
+
+    *s = 0.125f * hsum_float_8(accumf);
+
+#else
+
+    float sumf = 0.f;
+    for (int i = 0; i < nb; ++i) {
+        const float d = GGML_FP16_TO_FP32(x[i].d) * y[i].d;
+        const uint16_t * restrict q2 = x[i].qs;
+        const uint8_t  * restrict sc = x[i].scales;
+        const int8_t   * restrict q8 = y[i].qs;
+        int32_t bsum = 0;
+        for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+            const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
+            const uint16_t ls2 = 2*(sc[ib32] >>  4) + 1;
+            int32_t sumi = 0;
+            for (int l = 0; l < 2; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls1;
+            sumi = 0;
+            for (int l = 2; l < 4; ++l) {
+                const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+                const uint8_t  signs = ksigns_iq2xs[q2[l] >> 9];
+                for (int j = 0; j < 8; ++j) {
+                    sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+                }
+                q8 += 8;
+            }
+            bsum += sumi * ls2;
+            q2 += 4;
+        }
+        sumf += d * bsum;
+    }
+    *s = 0.125f * sumf;
+#endif
+}

ggml-quants.h

@@ -165,6 +165,22 @@ typedef struct {
 } block_q8_K;
 static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
 
+// (Almost) "true" 2-bit quantization.
+// Due to the need to use blocks as per ggml dsign, it ends up using
+// 2.0625 bpw because of the 16-bit scale for each block of 256.
+typedef struct {
+    ggml_fp16_t d;
+    uint16_t qs[QK_K/8];
+} block_iq2_xxs;
+static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
+
+// 2.3125 bpw quants
+typedef struct {
+    ggml_fp16_t d;
+    uint16_t qs[QK_K/8];
+    uint8_t  scales[QK_K/32];
+} block_iq2_xs;
+static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
 
 // Quantization
 void quantize_row_q4_0_reference(const float * restrict x, block_q4_0 * restrict y, int k);
@@ -180,6 +196,8 @@ 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);
@@ -194,6 +212,8 @@ 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);
@@ -209,6 +229,8 @@ void dequantize_row_q4_K(const block_q4_K * restrict x, float * restrict y, int
 void dequantize_row_q5_K(const block_q5_K * restrict x, float * restrict y, int k);
 void dequantize_row_q6_K(const block_q6_K * restrict x, float * restrict y, int k);
 void dequantize_row_q8_K(const block_q8_K * restrict x, float * restrict y, int k);
+void dequantize_row_iq2_xxs(const block_iq2_xxs * restrict x, float * restrict y, int k);
+void dequantize_row_iq2_xs (const block_iq2_xs  * restrict x, float * restrict y, int k);
 
 // Dot product
 void ggml_vec_dot_q4_0_q8_0(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
@@ -222,3 +244,5 @@ void ggml_vec_dot_q3_K_q8_K(int n, float * restrict s, const void * restrict vx,
 void ggml_vec_dot_q4_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 void ggml_vec_dot_q5_K_q8_K(int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 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);

ggml.c

@@ -132,7 +132,7 @@ void ggml_print_backtrace(void) {
             "-ex", "bt -frame-info source-and-location",
             "-ex", "detach",
             "-ex", "quit",
-            NULL);
+            (char *) NULL);
     } else {
         waitpid(pid, NULL, 0);
     }
@@ -394,6 +394,12 @@ static const size_t CACHE_LINE_SIZE_F32 = CACHE_LINE_SIZE/sizeof(float);
 static void ggml_vec_dot_f32(const int n, float * restrict s, const float * restrict x, const float * restrict y);
 static void ggml_vec_dot_f16(const int n, float * restrict s, ggml_fp16_t * restrict x, ggml_fp16_t * restrict y);
 
+ggml_collect_imatrix_t g_imatrix_collect = NULL;
+
+void ggml_set_imatrix_collection(ggml_collect_imatrix_t imatrix_collect) {
+    g_imatrix_collect = imatrix_collect;
+}
+
 static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
     [GGML_TYPE_I8] = {
         .type_name                = "i8",
@@ -573,6 +579,28 @@ static const ggml_type_traits_t type_traits[GGML_TYPE_COUNT] = {
         .vec_dot                  = ggml_vec_dot_q6_K_q8_K,
         .vec_dot_type             = GGML_TYPE_Q8_K,
     },
+    [GGML_TYPE_IQ2_XXS] = {
+        .type_name                = "iq2_xxs",
+        .blck_size                = QK_K,
+        .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,
+        .vec_dot                  = ggml_vec_dot_iq2_xxs_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+    },
+    [GGML_TYPE_IQ2_XS] = {
+        .type_name                = "iq2_xs",
+        .blck_size                = QK_K,
+        .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,
+        .vec_dot                  = ggml_vec_dot_iq2_xs_q8_K,
+        .vec_dot_type             = GGML_TYPE_Q8_K,
+    },
     [GGML_TYPE_Q8_K] = {
         .type_name                = "q8_K",
         .blck_size                = QK_K,
@@ -2111,6 +2139,8 @@ enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
         case GGML_FTYPE_MOSTLY_Q4_K:          wtype = GGML_TYPE_Q4_K;  break;
         case GGML_FTYPE_MOSTLY_Q5_K:          wtype = GGML_TYPE_Q5_K;  break;
         case GGML_FTYPE_MOSTLY_Q6_K:          wtype = GGML_TYPE_Q6_K;  break;
+        case GGML_FTYPE_MOSTLY_IQ2_XXS:       wtype = GGML_TYPE_IQ2_XXS;  break;
+        case GGML_FTYPE_MOSTLY_IQ2_XS:        wtype = GGML_TYPE_IQ2_XS;   break;
         case GGML_FTYPE_UNKNOWN:              wtype = GGML_TYPE_COUNT; break;
         case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
     }
@@ -4299,13 +4329,13 @@ struct ggml_tensor * ggml_set_2d_inplace(
 static struct ggml_tensor * ggml_cpy_impl(
         struct ggml_context * ctx,
         struct ggml_tensor  * a,
-        struct ggml_tensor  * b,
-        bool inplace) {
+        struct ggml_tensor  * b) {
     GGML_ASSERT(ggml_nelements(a) == ggml_nelements(b));
 
     bool is_node = false;
 
-    if (!inplace && (a->grad || b->grad)) {
+    if (a->grad || b->grad) {
+        // inplace is false and either one have a grad
         is_node = true;
     }
 
@@ -4329,29 +4359,21 @@ struct ggml_tensor * ggml_cpy(
         struct ggml_context * ctx,
         struct ggml_tensor * a,
         struct ggml_tensor * b) {
-    return ggml_cpy_impl(ctx, a, b, false);
-}
-
-struct ggml_tensor * ggml_cpy_inplace(
-        struct ggml_context * ctx,
-        struct ggml_tensor * a,
-        struct ggml_tensor * b) {
-    return ggml_cpy_impl(ctx, a, b, true);
+    return ggml_cpy_impl(ctx, a, b);
 }
 
 // ggml_cont
 
 static struct ggml_tensor * ggml_cont_impl(
         struct ggml_context * ctx,
-        struct ggml_tensor  * a,
-        bool inplace) {
+        struct ggml_tensor  * a) {
     bool is_node = false;
 
-    if (!inplace && a->grad) {
+    if (a->grad) {
         is_node = true;
     }
 
-    struct ggml_tensor * result = inplace ? ggml_view_tensor(ctx, a) : ggml_dup_tensor(ctx, a);
+    struct ggml_tensor * result = ggml_dup_tensor(ctx, a);
     ggml_format_name(result, "%s (cont)", a->name);
 
     result->op   = GGML_OP_CONT;
@@ -4364,13 +4386,7 @@ static struct ggml_tensor * ggml_cont_impl(
 struct ggml_tensor * ggml_cont(
         struct ggml_context * ctx,
         struct ggml_tensor * a) {
-    return ggml_cont_impl(ctx, a, false);
-}
-
-struct ggml_tensor * ggml_cont_inplace(
-        struct ggml_context * ctx,
-        struct ggml_tensor * a) {
-    return ggml_cont_impl(ctx, a, true);
+    return ggml_cont_impl(ctx, a);
 }
 
 // make contiguous, with new shape
@@ -7436,6 +7452,8 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
             {
                 ggml_compute_forward_add_q_f32(params, src0, src1, dst);
             } break;
@@ -7700,6 +7718,8 @@ static void ggml_compute_forward_add1(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
             {
                 ggml_compute_forward_add1_q_f32(params, src0, src1, dst);
             } break;
@@ -7814,6 +7834,8 @@ static void ggml_compute_forward_acc(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
         default:
             {
                 GGML_ASSERT(false);
@@ -9747,6 +9769,10 @@ static void ggml_compute_forward_mul_mat(
     const int ith = params->ith;
     const int nth = params->nth;
 
+    if (ith == 1 && g_imatrix_collect) {
+        g_imatrix_collect(src0, src1);
+    }
+
     const enum ggml_type type = src0->type;
 
     const bool src1_cont = ggml_is_contiguous(src1);
@@ -10050,6 +10076,10 @@ static void ggml_compute_forward_mul_mat_id(
 
         const struct ggml_tensor * src0_cur = dst->src[cur_a + 2];
 
+        if (ith == 1 && g_imatrix_collect) {
+            g_imatrix_collect(src0_cur, src1);
+        }
+
         const void * wdata    = (src1->type == vec_dot_type) ? src1->data : params->wdata;
         const size_t row_size = ggml_row_size(vec_dot_type, ne10);
 
@@ -10455,6 +10485,8 @@ static void ggml_compute_forward_out_prod(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
             {
                 ggml_compute_forward_out_prod_q_f32(params, src0, src1, dst);
             } break;
@@ -10629,6 +10661,8 @@ static void ggml_compute_forward_set(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
         default:
             {
                 GGML_ASSERT(false);
@@ -10823,6 +10857,8 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
             {
                 ggml_compute_forward_get_rows_q(params, src0, src1, dst);
             } break;
@@ -11459,6 +11495,8 @@ static void ggml_compute_forward_alibi(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_Q8_K:
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
@@ -11533,6 +11571,8 @@ static void ggml_compute_forward_clamp(
         case GGML_TYPE_Q4_K:
         case GGML_TYPE_Q5_K:
         case GGML_TYPE_Q6_K:
+        case GGML_TYPE_IQ2_XXS:
+        case GGML_TYPE_IQ2_XS:
         case GGML_TYPE_Q8_K:
         case GGML_TYPE_I8:
         case GGML_TYPE_I16:
@@ -18648,6 +18688,18 @@ size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, i
                 block_q6_K * block = (block_q6_K*)dst + start / QK_K;
                 result = ggml_quantize_q6_K(src + start, block, n, n, hist);
             } break;
+        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);
+            } 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);
+            } break;
         case GGML_TYPE_F16:
             {
                 int elemsize = sizeof(ggml_fp16_t);
@@ -19003,8 +19055,8 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
                 (int64_t) info->ne[3];
 
             if (ne % ggml_blck_size(info->type) != 0) {
-                fprintf(stderr, "%s: tensor '%s' number of elements (%" PRId64 ") is not a multiple of block size (%d)\n",
-                        __func__, info->name.data, ne, ggml_blck_size(info->type));
+                fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%d)\n",
+                        __func__, info->name.data, (int)info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type));
                 fclose(file);
                 gguf_free(ctx);
                 return NULL;

ggml.h

@@ -218,7 +218,9 @@
 #define GGML_MAX_PARAMS         2048
 #define GGML_MAX_CONTEXTS       64
 #define GGML_MAX_SRC            10
+#ifndef GGML_MAX_NAME
 #define GGML_MAX_NAME           64
+#endif
 #define GGML_MAX_OP_PARAMS      64
 #define GGML_DEFAULT_N_THREADS  4
 #define GGML_DEFAULT_GRAPH_SIZE 2048
@@ -339,6 +341,8 @@ extern "C" {
         GGML_TYPE_Q5_K = 13,
         GGML_TYPE_Q6_K = 14,
         GGML_TYPE_Q8_K = 15,
+        GGML_TYPE_IQ2_XXS = 16,
+        GGML_TYPE_IQ2_XS  = 17,
         GGML_TYPE_I8,
         GGML_TYPE_I16,
         GGML_TYPE_I32,
@@ -373,6 +377,8 @@ extern "C" {
         GGML_FTYPE_MOSTLY_Q4_K = 12, // except 1d tensors
         GGML_FTYPE_MOSTLY_Q5_K = 13, // except 1d tensors
         GGML_FTYPE_MOSTLY_Q6_K = 14, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ2_XXS = 15, // except 1d tensors
+        GGML_FTYPE_MOSTLY_IQ2_XS  = 16, // except 1d tensors
     };
 
     // available tensor operations:
@@ -1159,22 +1165,11 @@ extern "C" {
             struct ggml_tensor  * a,
             struct ggml_tensor  * b);
 
-    // a -> b, in-place, return view(b)
-    GGML_API struct ggml_tensor * ggml_cpy_inplace(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a,
-            struct ggml_tensor  * b);
-
     // make contiguous
     GGML_API struct ggml_tensor * ggml_cont(
             struct ggml_context * ctx,
             struct ggml_tensor  * a);
 
-    // make contiguous, in-place
-    GGML_API struct ggml_tensor * ggml_cont_inplace(
-            struct ggml_context * ctx,
-            struct ggml_tensor  * a);
-
     // make contiguous, with new shape
     GGML_API struct ggml_tensor * ggml_cont_1d(
             struct ggml_context * ctx,
@@ -2067,9 +2062,17 @@ 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);
 
+    //
+    // Importance matrix
+    //
+    typedef void(*ggml_collect_imatrix_t)(const struct ggml_tensor * src0, const struct ggml_tensor * src1);
+    GGML_API void ggml_set_imatrix_collection(ggml_collect_imatrix_t imatrix_collect);
+
     //
     // gguf
     //

gguf-py/gguf/tensor_mapping.py

@@ -57,6 +57,7 @@ class TensorNameMap:
             "transformer.norm_f",                      # mpt
             "ln_f",                                    # refact bloom qwen gpt2
             "language_model.encoder.final_layernorm",  # persimmon
+            "model.final_layernorm",                   # persimmon
             "lm_head.ln",                              # phi2
         ),
 
@@ -98,6 +99,7 @@ class TensorNameMap:
             "transformer.h.{bid}.self_attention.query_key_value",                  # falcon
             "h.{bid}.self_attention.query_key_value",                              # bloom
             "language_model.encoder.layers.{bid}.self_attention.query_key_value",  # persimmon
+            "model.layers.{bid}.self_attn.query_key_value",                        # persimmon
             "h.{bid}.attn.c_attn",                                                 # gpt2
             "transformer.h.{bid}.mixer.Wqkv",                                      # phi2
         ),
@@ -141,6 +143,7 @@ class TensorNameMap:
             "encoder.layer.{bid}.attention.output.dense",                # bert
             "transformer.h.{bid}.attn.out_proj",                         # gpt-j
             "language_model.encoder.layers.{bid}.self_attention.dense",  # persimmon
+            "model.layers.{bid}.self_attn.dense",                        # persimmon
             "h.{bid}.attn.c_proj",                                       # gpt2
             "transformer.h.{bid}.mixer.out_proj",                        # phi2
             "model.layers.layers.{bid}.self_attn.o_proj",                # plamo
@@ -184,6 +187,7 @@ class TensorNameMap:
             "encoder.layer.{bid}.intermediate.dense",                 # bert
             "transformer.h.{bid}.mlp.fc_in",                          # gpt-j
             "language_model.encoder.layers.{bid}.mlp.dense_h_to_4h",  # persimmon
+            "model.layers.{bid}.mlp.dense_h_to_4h",                   # persimmon
             "transformer.h.{bid}.mlp.w1",                             # qwen
             "h.{bid}.mlp.c_fc",                                       # gpt2
             "transformer.h.{bid}.mlp.fc1",                            # phi2
@@ -225,6 +229,7 @@ class TensorNameMap:
             "encoder.layer.{bid}.output.dense",                       # bert
             "transformer.h.{bid}.mlp.fc_out",                         # gpt-j
             "language_model.encoder.layers.{bid}.mlp.dense_4h_to_h",  # persimmon
+            "model.layers.{bid}.mlp.dense_4h_to_h",                   # persimmon
             "h.{bid}.mlp.c_proj",                                     # gpt2
             "transformer.h.{bid}.mlp.fc2",                            # phi2
             "model.layers.layers.{bid}.mlp.down_proj",                # plamo
@@ -237,10 +242,12 @@ class TensorNameMap:
 
         MODEL_TENSOR.ATTN_Q_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.q_layernorm",
+            "model.layers.{bid}.self_attn.q_layernorm",                       # persimmon
         ),
 
         MODEL_TENSOR.ATTN_K_NORM: (
             "language_model.encoder.layers.{bid}.self_attention.k_layernorm",
+            "model.layers.{bid}.self_attn.k_layernorm",                       # persimmon
         ),
 
         MODEL_TENSOR.ROPE_FREQS: (

llama.cpp

@@ -1903,6 +1903,28 @@ static void llama_kv_cache_seq_shift(
     cache.head = new_head != cache.size ? new_head : 0;
 }
 
+static void llama_kv_cache_seq_div(
+        struct llama_kv_cache & cache,
+                 llama_seq_id   seq_id,
+                    llama_pos   p0,
+                    llama_pos   p1,
+                          int   d) {
+    if (p0 < 0) p0 = 0;
+    if (p1 < 0) p1 = std::numeric_limits<llama_pos>::max();
+
+    for (uint32_t i = 0; i < cache.size; ++i) {
+        if (cache.cells[i].has_seq_id(seq_id) && cache.cells[i].pos >= p0 && cache.cells[i].pos < p1) {
+            cache.has_shift = true;
+
+            {
+                llama_pos p_old = cache.cells[i].pos;
+                cache.cells[i].pos   /= d;
+                cache.cells[i].delta += cache.cells[i].pos - p_old;
+            }
+        }
+    }
+}
+
 //
 // model loading and saving
 //
@@ -2200,6 +2222,8 @@ struct llama_model_loader {
                 case GGML_TYPE_Q4_K: ftype = LLAMA_FTYPE_MOSTLY_Q4_K_M; break;
                 case GGML_TYPE_Q5_K: ftype = LLAMA_FTYPE_MOSTLY_Q5_K_M; break;
                 case GGML_TYPE_Q6_K: ftype = LLAMA_FTYPE_MOSTLY_Q6_K;   break;
+                case GGML_TYPE_IQ2_XXS: ftype = LLAMA_FTYPE_MOSTLY_IQ2_XXS; break;
+                case GGML_TYPE_IQ2_XS:  ftype = LLAMA_FTYPE_MOSTLY_IQ2_XS;  break;
                 default:
                     {
                         LLAMA_LOG_WARN("%s: unknown type %s\n", __func__, ggml_type_name(type_max));
@@ -2562,7 +2586,8 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q8_0: return "Q8_0";
 
         // K-quants
-        case LLAMA_FTYPE_MOSTLY_Q2_K:   return "Q2_K";
+        case LLAMA_FTYPE_MOSTLY_Q2_K:   return "Q2_K - Medium";
+        case LLAMA_FTYPE_MOSTLY_Q2_K_S: return "Q2_K - Small";
         case LLAMA_FTYPE_MOSTLY_Q3_K_S: return "Q3_K - Small";
         case LLAMA_FTYPE_MOSTLY_Q3_K_M: return "Q3_K - Medium";
         case LLAMA_FTYPE_MOSTLY_Q3_K_L: return "Q3_K - Large";
@@ -2571,6 +2596,8 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q5_K_S: return "Q5_K - Small";
         case LLAMA_FTYPE_MOSTLY_Q5_K_M: return "Q5_K - Medium";
         case LLAMA_FTYPE_MOSTLY_Q6_K:   return "Q6_K";
+        case LLAMA_FTYPE_MOSTLY_IQ2_XXS:return "IQ2_XSS - 2.0625 bpw";
+        case LLAMA_FTYPE_MOSTLY_IQ2_XS: return "IQ2_XS - 2.3125 bpw";
 
         default: return "unknown, may not work";
     }
@@ -2805,6 +2832,7 @@ static void llm_load_hparams(
                 ml.get_key(LLM_KV_ATTENTION_LAYERNORM_EPS, hparams.f_norm_eps);
 
                 switch (hparams.n_layer) {
+                    case 24: model.type = e_model::MODEL_1B; break;
                     case 32: model.type = e_model::MODEL_3B; break;
                     default: model.type = e_model::MODEL_UNKNOWN;
                 }
@@ -3121,7 +3149,15 @@ static void llm_load_print_meta(llama_model_loader & ml, llama_model & model) {
     LLAMA_LOG_INFO("%s: rope_finetuned   = %s\n",     __func__, hparams.rope_finetuned ? "yes" : "unknown");
     LLAMA_LOG_INFO("%s: model type       = %s\n",     __func__, llama_model_type_name(model.type));
     LLAMA_LOG_INFO("%s: model ftype      = %s\n",     __func__, llama_model_ftype_name(model.ftype).c_str());
-    LLAMA_LOG_INFO("%s: model params     = %.2f B\n", __func__, ml.n_elements*1e-9);
+    if (ml.n_elements >= 1e12) {
+        LLAMA_LOG_INFO("%s: model params     = %.2f T\n", __func__, ml.n_elements*1e-12);
+    } else if (ml.n_elements >= 1e9) {
+        LLAMA_LOG_INFO("%s: model params     = %.2f B\n", __func__, ml.n_elements*1e-9);
+    } else if (ml.n_elements >= 1e6) {
+        LLAMA_LOG_INFO("%s: model params     = %.2f M\n", __func__, ml.n_elements*1e-6);
+    } else {
+        LLAMA_LOG_INFO("%s: model params     = %.2f K\n", __func__, ml.n_elements*1e-3);
+    }
     if (ml.n_bytes < GiB) {
         LLAMA_LOG_INFO("%s: model size       = %.2f MiB (%.2f BPW) \n", __func__, ml.n_bytes/1024.0/1024.0,        ml.n_bytes*8.0/ml.n_elements);
     } else {
@@ -4068,7 +4104,6 @@ static void llm_build_k_shift(
        struct ggml_cgraph * graph,
             llm_rope_type   type,
                   int64_t   n_ctx,
-                  int       n_rot,
                   float     freq_base,
                   float     freq_scale,
        const llm_build_cb & cb) {
@@ -4076,14 +4111,13 @@ static void llm_build_k_shift(
     const int64_t n_head_kv     = hparams.n_head_kv;
     const int64_t n_embd_head_k = hparams.n_embd_head_k;
     const int64_t n_embd_k_gqa  = hparams.n_embd_k_gqa();
+    const int32_t n_rot         = hparams.n_rot;
     const int32_t n_orig_ctx    = cparams.n_yarn_orig_ctx;
     const float   ext_factor    = cparams.yarn_ext_factor;
     const float   attn_factor   = cparams.yarn_attn_factor;
     const float   beta_fast     = cparams.yarn_beta_fast;
     const float   beta_slow     = cparams.yarn_beta_slow;
 
-    GGML_ASSERT(n_embd_head_k % n_rot == 0);
-
     struct ggml_tensor * K_shift = ggml_new_tensor_1d(ctx, GGML_TYPE_I32, n_ctx);
     cb(K_shift, "K_shift", -1);
 
@@ -4487,7 +4521,7 @@ struct llm_build_context {
 
         // shift the entire K-cache if needed
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -4525,14 +4559,14 @@ struct llm_build_context {
 
                 Qcur = ggml_rope_custom(
                     ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
-                    n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale,
+                    hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Qcur, "Qcur", il);
 
                 Kcur = ggml_rope_custom(
                     ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                    n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale,
+                    hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
                     ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -4655,6 +4689,7 @@ struct llm_build_context {
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -4672,7 +4707,7 @@ struct llm_build_context {
 
         // shift the entire K-cache if needed
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -4698,12 +4733,12 @@ struct llm_build_context {
                     case MODEL_7B:
                         Qcur = ggml_rope_custom(
                             ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head, n_tokens), inp_pos,
-                            n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale,
+                            hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
                             ext_factor, attn_factor, beta_fast, beta_slow
                         );
                         Kcur = ggml_rope_custom(
                             ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
-                            n_embd_head, 0, 0, n_orig_ctx, freq_base, freq_scale,
+                            hparams.n_rot, 0, 0, n_orig_ctx, freq_base, freq_scale,
                             ext_factor, attn_factor, beta_fast, beta_slow
                         );
                         break;
@@ -4776,6 +4811,7 @@ struct llm_build_context {
         const int64_t n_embd_head = hparams.n_embd_head_v;
         const int64_t n_embd_gqa  = hparams.n_embd_v_gqa();
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -4793,7 +4829,7 @@ struct llm_build_context {
 
         // shift the entire K-cache if needed
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -4834,13 +4870,13 @@ struct llm_build_context {
 
                 // using mode = 2 for neox mode
                 Qcur = ggml_rope_custom(
-                    ctx0, Qcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx,
+                    ctx0, Qcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Qcur, "Qcur", il);
 
                 Kcur = ggml_rope_custom(
-                    ctx0, Kcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx,
+                    ctx0, Kcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -4997,9 +5033,8 @@ struct llm_build_context {
         struct ggml_cgraph * gf = ggml_new_graph_custom(ctx0, LLAMA_MAX_NODES, false);
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
-        GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
-
-        const int64_t n_rot = n_embd_head_k / 2;
+        GGML_ASSERT(n_embd_head   == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head/2 == hparams.n_rot);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -5016,7 +5051,7 @@ struct llm_build_context {
         cb(KQ_mask, "KQ_mask", -1);
 
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -5076,7 +5111,7 @@ struct llm_build_context {
 
                 // RoPE the first n_rot of q/k, pass the other half, and concat.
                 struct ggml_tensor * qrot = ggml_view_3d(
-                        ctx0, tmpq, n_rot, n_head, n_tokens,
+                        ctx0, tmpq, hparams.n_rot, n_head, n_tokens,
                         ggml_element_size(tmpq) * n_embd_head,
                         ggml_element_size(tmpq) * n_embd_head * n_head,
                         0
@@ -5084,7 +5119,7 @@ struct llm_build_context {
                 cb(qrot, "qrot", il);
 
                 struct ggml_tensor * krot = ggml_view_3d(
-                        ctx0, tmpk, n_rot, n_head, n_tokens,
+                        ctx0, tmpk, hparams.n_rot, n_head, n_tokens,
                         ggml_element_size(tmpk) * n_embd_head,
                         ggml_element_size(tmpk) * n_embd_head * n_head,
                         0
@@ -5093,29 +5128,29 @@ struct llm_build_context {
 
                 // get the second half of tmpq, e.g tmpq[n_rot:, :, :]
                 struct ggml_tensor * qpass = ggml_view_3d(
-                        ctx0, tmpq, n_rot, n_head, n_tokens,
+                        ctx0, tmpq, hparams.n_rot, n_head, n_tokens,
                         ggml_element_size(tmpq) * n_embd_head,
                         ggml_element_size(tmpq) * n_embd_head * n_head,
-                        ggml_element_size(tmpq) * n_rot
+                        ggml_element_size(tmpq) * hparams.n_rot
                         );
                 cb(qpass, "qpass", il);
 
                 struct ggml_tensor * kpass = ggml_view_3d(
-                        ctx0, tmpk, n_rot, n_head, n_tokens,
+                        ctx0, tmpk, hparams.n_rot, n_head, n_tokens,
                         ggml_element_size(tmpk) * n_embd_head,
                         ggml_element_size(tmpk) * n_embd_head * n_head,
-                        ggml_element_size(tmpk) * n_rot
+                        ggml_element_size(tmpk) * hparams.n_rot
                         );
                 cb(kpass, "kpass", il);
 
                 struct ggml_tensor * qrotated = ggml_rope_custom(
-                    ctx0, qrot, inp_pos, n_rot, 2, 0, n_orig_ctx,
+                    ctx0, qrot, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(qrotated, "qrotated", il);
 
                 struct ggml_tensor * krotated = ggml_rope_custom(
-                    ctx0, krot, inp_pos, n_rot, 2, 0, n_orig_ctx,
+                    ctx0, krot, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(krotated, "krotated", il);
@@ -5495,6 +5530,7 @@ struct llm_build_context {
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -5512,7 +5548,7 @@ struct llm_build_context {
 
         // shift the entire K-cache if needed
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, hparams.n_rot, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -5625,7 +5661,7 @@ struct llm_build_context {
 
         // shift the entire K-cache if needed
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -5657,13 +5693,13 @@ struct llm_build_context {
 
                 // using mode = 2 for neox mode
                 Qcur = ggml_rope_custom(
-                    ctx0, Qcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx,
+                    ctx0, Qcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Qcur, "Qcur", il);
 
                 Kcur = ggml_rope_custom(
-                    ctx0, Kcur, inp_pos, n_embd_head, 2, 0, n_orig_ctx,
+                    ctx0, Kcur, inp_pos, hparams.n_rot, 2, 0, n_orig_ctx,
                     freq_base, freq_scale, ext_factor, attn_factor, beta_fast, beta_slow
                 );
                 cb(Kcur, "Kcur", il);
@@ -5742,7 +5778,7 @@ struct llm_build_context {
 
         // shift the entire K-cache if needed
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, n_embd_head, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE_NEOX, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -5838,6 +5874,7 @@ struct llm_build_context {
 
         const int64_t n_embd_head = hparams.n_embd_head_v;
         GGML_ASSERT(n_embd_head == hparams.n_embd_head_k);
+        GGML_ASSERT(n_embd_head == hparams.n_rot);
 
         struct ggml_tensor * cur;
         struct ggml_tensor * inpL;
@@ -5855,7 +5892,7 @@ struct llm_build_context {
 
         // shift the entire K-cache if needed
         if (do_rope_shift) {
-            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, n_embd_head, freq_base, freq_scale, cb);
+            llm_build_k_shift(ctx0, hparams, cparams, kv_self, gf, LLM_ROPE, n_ctx, freq_base, freq_scale, cb);
         }
 
         for (int il = 0; il < n_layer; ++il) {
@@ -5881,13 +5918,13 @@ struct llm_build_context {
                 cb(Vcur, "Vcur", il);
 
                 Qcur = ggml_rope_custom(
-                        ctx0, ggml_reshape_3d(ctx0, Qcur, n_embd_head, n_head,    n_tokens), inp_pos,
+                        ctx0, ggml_reshape_3d(ctx0, Qcur, hparams.n_rot, n_head,    n_tokens), inp_pos,
                         n_embd_head, 2, 0, n_orig_ctx, freq_base, freq_scale,
                         ext_factor, attn_factor, beta_fast, beta_slow);
                 cb(Qcur, "Qcur", il);
 
                 Kcur = ggml_rope_custom(
-                        ctx0, ggml_reshape_3d(ctx0, Kcur, n_embd_head, n_head_kv, n_tokens), inp_pos,
+                        ctx0, ggml_reshape_3d(ctx0, Kcur, hparams.n_rot, n_head_kv, n_tokens), inp_pos,
                         n_embd_head, 2, 0, n_orig_ctx, freq_base, freq_scale,
                         ext_factor, attn_factor, beta_fast, beta_slow);
                 cb(Kcur, "Kcur", il);
@@ -8920,10 +8957,13 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
             // TODO: explore better strategies
             new_type = GGML_TYPE_Q8_0;
         }
-    } else if (name.find("ffn_down.weight") != std::string::npos) {
+    } else if (name.find("ffn_down") != std::string::npos) {
         if      (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K_S) {
+            if (qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) new_type = GGML_TYPE_Q4_K;
+        }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
-            new_type = qs.i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K
+            new_type = qs.i_feed_forward_w2 < qs.n_feed_forward_w2/16 ? GGML_TYPE_Q5_K
                      : arch != LLM_ARCH_FALCON || use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q4_K
                      : GGML_TYPE_Q3_K;
         }
@@ -8932,14 +8972,14 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
             if (arch == LLM_ARCH_FALCON) {
-                new_type = qs.i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K :
+                new_type = qs.i_feed_forward_w2 < qs.n_feed_forward_w2/16 ? GGML_TYPE_Q6_K :
                            use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
             } else {
                 if (use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
             }
         }
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(qs.i_feed_forward_w2, qs.n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
-        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && qs.i_feed_forward_w2 < 4) {
+        else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && arch != LLM_ARCH_FALCON && qs.i_feed_forward_w2 < qs.n_feed_forward_w2/8) {
             new_type = GGML_TYPE_Q5_K;
         }
         ++qs.i_feed_forward_w2;
@@ -8957,9 +8997,10 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K;
         else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
     }
-    else if (name.find("ffn_gate.weight") != std::string::npos || name.find("ffn_up.weight") != std::string::npos) {
-        if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
-    }
+    // IK: let's remove this, else Q2_K is almost the same as Q3_K_S
+    //else if (name.find("ffn_gate") != std::string::npos || name.find("ffn_up") != std::string::npos) {
+    //    if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
+    //}
     // This can be used to reduce the size of the Q5_K_S model.
     // The associated PPL increase is fully in line with the size reduction
     //else {
@@ -9008,6 +9049,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
 
         // K-quants
         case LLAMA_FTYPE_MOSTLY_Q2_K:   quantized_type = GGML_TYPE_Q2_K; break;
+        case LLAMA_FTYPE_MOSTLY_Q2_K_S: quantized_type = GGML_TYPE_Q2_K; break;
         case LLAMA_FTYPE_MOSTLY_Q3_K_S:
         case LLAMA_FTYPE_MOSTLY_Q3_K_M:
         case LLAMA_FTYPE_MOSTLY_Q3_K_L: quantized_type = GGML_TYPE_Q3_K; break;
@@ -9016,6 +9058,8 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         case LLAMA_FTYPE_MOSTLY_Q5_K_S:
         case LLAMA_FTYPE_MOSTLY_Q5_K_M: quantized_type = GGML_TYPE_Q5_K; break;
         case LLAMA_FTYPE_MOSTLY_Q6_K:   quantized_type = GGML_TYPE_Q6_K; break;
+        case LLAMA_FTYPE_MOSTLY_IQ2_XXS:quantized_type = GGML_TYPE_IQ2_XXS; break;
+        case LLAMA_FTYPE_MOSTLY_IQ2_XS :quantized_type = GGML_TYPE_IQ2_XS;  break;
 
         default: throw std::runtime_error(format("invalid output file type %d\n", ftype));
     }
@@ -9064,7 +9108,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         if (name.find("attn_v.weight") != std::string::npos || name.find("attn_qkv.weight") != std::string::npos) {
             ++qs.n_attention_wv;
         }
-        else if (name.find("ffn_down.weight") != std::string::npos) {
+        else if (name.find("ffn_down") != std::string::npos) {
             ++qs.n_feed_forward_w2;
         }
     }
@@ -10140,9 +10184,21 @@ void llama_kv_cache_seq_keep(struct llama_context * ctx, llama_seq_id seq_id) {
 }
 
 void llama_kv_cache_seq_shift(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, llama_pos delta) {
+    if (delta == 0) {
+        return;
+    }
+
     llama_kv_cache_seq_shift(ctx->kv_self, seq_id, p0, p1, delta);
 }
 
+void llama_kv_cache_seq_div(struct llama_context * ctx, llama_seq_id seq_id, llama_pos p0, llama_pos p1, int d) {
+    if (d == 1) {
+        return;
+    }
+
+    llama_kv_cache_seq_div(ctx->kv_self, seq_id, p0, p1, d);
+}
+
 // Returns the *maximum* size of the state
 size_t llama_get_state_size(const struct llama_context * ctx) {
     // we don't know size of rng until we actually serialize it. so reserve more than enough memory for its serialized state.
@@ -10875,7 +10931,7 @@ void llama_print_timings(struct llama_context * ctx) {
             __func__, timings.t_p_eval_ms, timings.n_p_eval, timings.t_p_eval_ms / timings.n_p_eval, 1e3 / timings.t_p_eval_ms * timings.n_p_eval);
     LLAMA_LOG_INFO("%s:        eval time = %10.2f ms / %5d runs   (%8.2f ms per token, %8.2f tokens per second)\n",
             __func__, timings.t_eval_ms, timings.n_eval, timings.t_eval_ms / timings.n_eval, 1e3 / timings.t_eval_ms * timings.n_eval);
-    LLAMA_LOG_INFO("%s:       total time = %10.2f ms\n", __func__, (timings.t_end_ms - timings.t_start_ms));
+    LLAMA_LOG_INFO("%s:       total time = %10.2f ms / %5d tokens\n", __func__, (timings.t_end_ms - timings.t_start_ms), (timings.n_p_eval + timings.n_eval));
 }
 
 void llama_reset_timings(struct llama_context * ctx) {

llama.h

@@ -103,6 +103,9 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q5_K_S        = 16, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q5_K_M        = 17, // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q6_K          = 18, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ2_XXS       = 19, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_IQ2_XS        = 20, // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q2_K_S        = 21, // except 1d tensors
 
         LLAMA_FTYPE_GUESSED = 1024, // not specified in the model file
     };
@@ -484,6 +487,17 @@ extern "C" {
                        llama_pos   p1,
                        llama_pos   delta);
 
+    // Integer division of the positions by factor of `d > 1`
+    // If the KV cache is RoPEd, the KV data is updated accordingly
+    // p0 < 0 : [0,  p1]
+    // p1 < 0 : [p0, inf)
+    LLAMA_API void llama_kv_cache_seq_div(
+            struct llama_context * ctx,
+                    llama_seq_id   seq_id,
+                       llama_pos   p0,
+                       llama_pos   p1,
+                             int   d);
+
     //
     // State / sessions
     //

scripts/compare-llama-bench.py

@@ -0,0 +1,356 @@
+#!/usr/bin/env python3
+
+import argparse
+import heapq
+import sys
+import os
+from glob import glob
+import sqlite3
+
+try:
+    import git
+    from tabulate import tabulate
+except ImportError:
+    print("ERROR: the following Python libraries are required: GitPython, tabulate.")
+    sys.exit(1)
+
+# 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"
+]
+
+# Properties that are boolean and are converted to Yes/No for the table:
+BOOL_PROPERTIES = ["cuda", "opencl", "metal", "gpu_blas", "blas"]
+
+# Header names for the table:
+PRETTY_NAMES = {
+    "cuda": "CUDA", "opencl": "OpenCL", "metal": "Metal", "gpu_blas": "GPU BLAS", "blas": "BLAS",
+    "cpu_info": "CPU", "gpu_info": "GPU", "model_filename": "File", "model_type": "Model",
+    "model_size": "Model Size [GiB]", "model_n_params": "Num. of Parameters",
+    "n_batch": "Batch size", "n_threads": "Threads", "type_k": "K type", "type_v": "V type",
+    "n_gpu_layers": "GPU layers", "main_gpu": "Main GPU", "no_kv_offload": "NKVO",
+    "mul_mat_q": "MMQ", "tensor_split": "Tensor split"
+}
+
+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.
+
+DESCRIPTION = """Creates tables from llama-bench data written to an SQLite database. Example usage (Linux):
+
+$ git checkout master
+$ make clean && make llama-bench
+$ ./llama-bench -o sql | sqlite3 llama-bench.sqlite
+$ git checkout some_branch
+$ make clean && make llama-bench
+$ ./llama-bench -o sql | sqlite3 llama-bench.sqlite
+$ ./scripts/compare-llama-bench.py
+
+Performance numbers from multiple runs per commit are averaged WITHOUT being weighted by the --repetitions parameter of llama-bench.
+"""
+
+parser = argparse.ArgumentParser(
+    description=DESCRIPTION, formatter_class=argparse.RawDescriptionHelpFormatter)
+help_b = (
+    "The baseline commit to compare performance to. "
+    "Accepts either a branch name, tag name, or commit hash. "
+    "Defaults to latest master commit with data."
+)
+parser.add_argument("-b", "--baseline", help=help_b)
+help_c = (
+    "The commit whose performance is to be compared to the baseline. "
+    "Accepts either a branch name, tag name, or commit hash. "
+    "Defaults to the non-master commit for which llama-bench was run most recently."
+)
+parser.add_argument("-c", "--compare", help=help_c)
+help_i = (
+    "Input SQLite file for comparing commits. "
+    "Defaults to 'llama-bench.sqlite' in the current working directory. "
+    "If no such file is found and there is exactly one .sqlite file in the current directory, "
+    "that file is instead used as input."
+)
+parser.add_argument("-i", "--input", help=help_i)
+help_o = (
+    "Output format for the table. "
+    "Defaults to 'pipe' (GitHub compatible). "
+    "Also supports e.g. 'latex' or 'mediawiki'. "
+    "See tabulate documentation for full list."
+)
+parser.add_argument("-o", "--output", help=help_o, default="pipe")
+help_s = (
+    "Columns to add to the table. "
+    "Accepts a comma-separated list of values. "
+    f"Legal values: {', '.join(KEY_PROPERTIES[:-2])}. "
+    "Defaults to model name (model_type) and CPU and/or GPU name (cpu_info, gpu_info) "
+    "plus any column where not all data points are the same. "
+    "If the columns are manually specified, then the results for each unique combination of the "
+    "specified values are averaged WITHOUT weighing by the --repetitions parameter of llama-bench."
+)
+parser.add_argument("-s", "--show", help=help_s)
+
+known_args, unknown_args = parser.parse_known_args()
+
+if unknown_args:
+    print(f"ERROR: Received unknown args: {unknown_args}.")
+    print()
+    parser.print_help()
+    sys.exit(1)
+
+input_file = known_args.input
+if input_file is None and os.path.exists("./llama-bench.sqlite"):
+    input_file = "llama-bench.sqlite"
+if input_file is None:
+    sqlite_files = glob("*.sqlite")
+    if len(sqlite_files) == 1:
+        input_file = sqlite_files[0]
+
+if input_file is None:
+    print("ERROR: Cannot find a suitable input file, please provide one.")
+    print()
+    parser.print_help()
+    sys.exit(1)
+
+connection = sqlite3.connect(input_file)
+cursor = connection.cursor()
+builds = cursor.execute("SELECT DISTINCT build_commit FROM test;").fetchall()
+
+try:
+    repo = git.Repo(".", search_parent_directories=True)
+except git.exc.InvalidGitRepositoryError:
+    repo = None
+
+
+def find_parent_in_data(commit):
+    """Helper function to find the most recent parent measured in number of commits for which there is data."""
+    heap = [(0, commit)]
+    seen_hexsha8 = set()
+    while heap:
+        depth, current_commit = heapq.heappop(heap)
+        current_hexsha8 = commit.hexsha[:8]
+        if (current_hexsha8,) in builds:
+            return current_hexsha8
+        for parent in commit.parents:
+            parent_hexsha8 = parent.hexsha[:8]
+            if parent_hexsha8 not in seen_hexsha8:
+                seen_hexsha8.add(parent_hexsha8)
+                heapq.heappush(heap, (depth + 1, parent))
+    return None
+
+
+def get_all_parent_hexsha8s(commit):
+    """Helper function to recursively get hexsha8 values for all parents of a commit."""
+    unvisited = [commit]
+    visited   = []
+
+    while unvisited:
+        current_commit = unvisited.pop(0)
+        visited.append(current_commit.hexsha[:8])
+        for parent in current_commit.parents:
+            if parent.hexsha[:8] not in visited:
+                unvisited.append(parent)
+
+    return visited
+
+
+def get_commit_name(hexsha8):
+    """Helper function to find a human-readable name for a commit if possible."""
+    if repo is None:
+        return hexsha8
+    for h in repo.heads:
+        if h.commit.hexsha[:8] == hexsha8:
+            return h.name
+    for t in repo.tags:
+        if t.commit.hexsha[:8] == hexsha8:
+            return t.name
+    return hexsha8
+
+
+def get_commit_hexsha8(name):
+    """Helper function to search for a commit given a human-readable name."""
+    if repo is None:
+        return None
+    for h in repo.heads:
+        if h.name == name:
+            return h.commit.hexsha[:8]
+    for t in repo.tags:
+        if t.name == name:
+            return t.commit.hexsha[:8]
+    return None
+
+
+hexsha8_baseline = name_baseline = None
+
+# If the user specified a baseline, try to find a commit for it:
+if known_args.baseline is not None:
+    if (known_args.baseline,) in builds:
+        hexsha8_baseline = known_args.baseline
+    if hexsha8_baseline is None:
+        hexsha8_baseline = get_commit_hexsha8(known_args.baseline)
+        name_baseline = known_args.baseline
+    if hexsha8_baseline is None:
+        print(f"ERROR: cannot find data for baseline={known_args.baseline}.")
+        sys.exit(1)
+# Otherwise, search for the most recent parent of master for which there is data:
+elif repo is not None:
+    hexsha8_baseline = find_parent_in_data(repo.heads.master.commit)
+
+    if hexsha8_baseline is None:
+        print("ERROR: No baseline was provided and did not find data for any master branch commits.")
+        print()
+        parser.print_help()
+        sys.exit(1)
+else:
+    print(
+        "ERROR: No baseline was provided and the current working directory "
+        "is not part of a git repository from which a baseline could be inferred."
+    )
+    print()
+    parser.print_help()
+    sys.exit(1)
+
+
+name_baseline = get_commit_name(hexsha8_baseline)
+
+hexsha8_compare = name_compare = None
+
+# If the user has specified a compare value, try to find a corresponding commit:
+if known_args.compare is not None:
+    if (known_args.compare,) in builds:
+        hexsha8_compare = known_args.compare
+    if hexsha8_compare is None:
+        hexsha8_compare = get_commit_hexsha8(known_args.compare)
+        name_compare = known_args.compare
+    if hexsha8_compare is None:
+        print(f"ERROR: cannot find data for baseline={known_args.compare}.")
+        sys.exit(1)
+# Otherwise, search for the commit for llama-bench was most recently run
+# and that is not a parent of master:
+elif repo is not None:
+    hexsha8s_master = get_all_parent_hexsha8s(repo.heads.master.commit)
+    builds_timestamp = cursor.execute(
+        "SELECT build_commit, test_time FROM test ORDER BY test_time;").fetchall()
+    for (hexsha8, _) in reversed(builds_timestamp):
+        if hexsha8 not in hexsha8s_master:
+            hexsha8_compare = hexsha8
+            break
+
+    if hexsha8_compare is None:
+        print("ERROR: No compare target was provided and did not find data for any non-master commits.")
+        print()
+        parser.print_help()
+        sys.exit(1)
+else:
+    print(
+        "ERROR: No compare target was provided and the current working directory "
+        "is not part of a git repository from which a compare target could be inferred."
+    )
+    print()
+    parser.print_help()
+    sys.exit(1)
+
+name_compare = get_commit_name(hexsha8_compare)
+
+
+def get_rows(properties):
+    """
+    Helper function that gets table rows for some list of properties.
+    Rows are created by combining those where all provided properties are equal.
+    The resulting rows are then grouped by the provided properties and the t/s values are averaged.
+    The returned rows are unique in terms of property combinations.
+    """
+    select_string = ", ".join(
+        [f"tb.{p}" for p in properties] + ["tb.n_prompt", "tb.n_gen", "AVG(tb.avg_ts)", "AVG(tc.avg_ts)"])
+    equal_string = " AND ".join(
+        [f"tb.{p} = tc.{p}" for p in KEY_PROPERTIES] + [
+            f"tb.build_commit = '{hexsha8_baseline}'", f"tc.build_commit = '{hexsha8_compare}'"]
+    )
+    group_order_string = ", ".join([f"tb.{p}" for p in properties] + ["tb.n_gen", "tb.n_prompt"])
+    query = (f"SELECT {select_string} FROM test tb JOIN test tc ON {equal_string} "
+             f"GROUP BY {group_order_string} ORDER BY {group_order_string};")
+    return cursor.execute(query).fetchall()
+
+
+# If the user provided columns to group the results by, use them:
+if known_args.show is not None:
+    show = known_args.show.split(",")
+    unknown_cols = []
+    for prop in show:
+        if prop not in KEY_PROPERTIES[:-2]:  # Last two values are n_prompt, n_gen.
+            unknown_cols.append(prop)
+    if unknown_cols:
+        print(f"ERROR: Unknown values for --show: {', '.join(unknown_cols)}")
+        print()
+        parser.print_usage()
+        sys.exit(1)
+    rows_show = get_rows(show)
+# Otherwise, select those columns where the values are not all the same:
+else:
+    rows_full = get_rows(KEY_PROPERTIES)
+    properties_different = []
+    for i, kp_i in enumerate(KEY_PROPERTIES):
+        if kp_i in DEFAULT_SHOW or kp_i == "n_prompt" or kp_i == "n_gen":
+            continue
+        for row_full in rows_full:
+            if row_full[i] != rows_full[0][i]:
+                properties_different.append(kp_i)
+                break
+
+    show = []
+    # Show CPU and/or GPU by default even if the hardware for all results is the same:
+    if "gpu_blas" not in properties_different and "n_gpu_layers" not in properties_different:
+        gpu_blas = bool(rows_full[0][KEY_PROPERTIES.index("gpu_blas")])
+        ngl = int(rows_full[0][KEY_PROPERTIES.index("n_gpu_layers")])
+
+        if not gpu_blas or ngl != 99 and "cpu_info" not in properties_different:
+            show.append("cpu_info")
+        if gpu_blas and "gpu_info" not in properties_different:
+            show.append("gpu_info")
+
+    show += DEFAULT_SHOW
+    show += properties_different
+    for prop in DEFAULT_HIDE:
+        try:
+            show.remove(prop)
+        except ValueError:
+            pass
+    rows_show = get_rows(show)
+
+table = []
+for row in rows_show:
+    n_prompt = int(row[-4])
+    n_gen    = int(row[-3])
+    assert n_prompt == 0 or n_gen == 0
+    test_name = f"tg{n_gen}" if n_prompt == 0 else f"pp{n_prompt}"
+    #           Regular columns    test name    avg t/s values              Speedup
+    #            VVVVVVVVVVVVV     VVVVVVVVV    VVVVVVVVVVVVVV              VVVVVVV
+    table.append(list(row[:-4]) + [test_name] + list(row[-2:]) + [float(row[-1]) / float(row[-2])])
+
+# Some a-posteriori fixes to make the table contents prettier:
+for bool_property in BOOL_PROPERTIES:
+    if bool_property in show:
+        ip = show.index(bool_property)
+        for row_table in table:
+            row_table[ip] = "Yes" if int(row_table[ip]) == 1 else "No"
+
+if "model_size" in show:
+    ip = show.index("model_size")
+    for row_table in table:
+        row_table[ip] = float(row_table[ip]) / 1024 ** 3
+
+if "gpu_info" in show:
+    ip = show.index("gpu_info")
+    for gns in GPU_NAME_STRIP:
+        for row_table in table:
+            row_table[ip] = row_table[ip].replace(gns, "")
+
+headers  = [PRETTY_NAMES[p] for p in show]
+headers += ["Test", f"t/s {name_baseline}", f"t/s {name_compare}", "Speedup"]
+
+print(tabulate(
+    table,
+    headers=headers,
+    floatfmt=".2f",
+    tablefmt=known_args.output
+))

scripts/get-pg.sh

@@ -0,0 +1,70 @@
+#!/bin/bash
+
+function usage {
+    echo "usage: <n>$0"
+    echo "note: n is the number of essays to download"
+    echo "for specific n, the resulting pg.txt file will have the following number of tokens:"
+    echo "n   | tokens"
+    echo "--- | ---"
+    echo "1   | 6230"
+    echo "2   | 23619"
+    echo "5   | 25859"
+    echo "10  | 36888"
+    echo "15  | 50188"
+    echo "20  | 59094"
+    echo "25  | 88764"
+    echo "30  | 103121"
+    echo "32  | 108338"
+    echo "35  | 113403"
+    echo "40  | 127699"
+    echo "45  | 135896"
+    exit 1
+}
+
+function has_cmd {
+    if ! [ -x "$(command -v $1)" ]; then
+        echo "error: $1 is not available" >&2
+        exit 1
+    fi
+}
+
+# check for: curl, html2text, tail, sed, fmt
+has_cmd curl
+has_cmd html2text
+has_cmd tail
+has_cmd sed
+
+if [ $# -ne 1 ]; then
+    usage
+fi
+
+n=$1
+
+# get urls
+urls="$(curl http://www.aaronsw.com/2002/feeds/pgessays.rss | grep html | sed -e "s/.*http/http/" | sed -e "s/html.*/html/" | head -n $n)"
+
+printf "urls:\n%s\n" "$urls"
+
+if [ -f pg.txt ]; then
+    rm pg.txt
+fi
+
+c=1
+for url in $urls; do
+    echo "processing $url"
+
+    cc=$(printf "%03d" $c)
+
+    curl -L $url | html2text | tail -n +4 | sed -E "s/^[[:space:]]+//g" | fmt -w 80 >> pg-$cc-one.txt
+    cat pg-$cc-one.txt >> pg.txt
+
+    cp -v pg.txt pg-$cc-all.txt
+    c=$((c+1))
+
+    # don't flood the server
+    sleep 1
+done
+
+echo "done. data in pg.txt"
+
+exit 0

scripts/sync-ggml.last

@@ -1 +1 @@
-f96711108d55bdbbd277e6be07204dce6a94fb93
+979cc23b345006504cfc1f67c0fdf627805e3319

spm-headers/ggml.h

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

tests/test-backend-ops.cpp

@@ -450,7 +450,7 @@ struct test_case {
 
             double err = nmse(f1.data(), f2.data(), f1.size());
             if (err > ud->max_err) {
-                printf("[%s] NMSE = %f ", ggml_op_desc(t1), err);
+                printf("[%s] NMSE = %.9f > %.9f ", ggml_op_desc(t1), err, ud->max_err);
                 //for (int i = 0; i < (int) f1.size(); i++) {
                 //    printf("%5d %9.6f %9.6f, diff = %9.6f\n", i, f1[i], f2[i], f1[i] - f2[i]);
                 //}
@@ -1449,6 +1449,7 @@ struct test_moe : public test_case {
 
 static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op_name) {
     std::vector<std::unique_ptr<test_case>> test_cases;
+    std::default_random_engine rng(0);
 
     const ggml_type all_types[] = {
         GGML_TYPE_F32, GGML_TYPE_F16,
@@ -1583,7 +1584,19 @@ static bool test_backend(ggml_backend_t backend, test_mode mode, const char * op
     test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 10,  1}, 5));
     test_cases.emplace_back(new test_diag_mask_inf(GGML_TYPE_F32, {10, 10, 10, 10}, 5));
 
-    test_cases.emplace_back(new test_soft_max());
+    std::uniform_int_distribution<> dist_ne1(1, 50);
+    int exponent = 1;
+    while (exponent < (1 << 17)) {
+        std::uniform_int_distribution<> dist_ne0(exponent, 2*exponent);
+
+        for (int n = 0; n < 10; ++n) {
+            int64_t ne0 = dist_ne0(rng);
+            int64_t ne1 = dist_ne1(rng);
+            test_cases.emplace_back(new test_soft_max(GGML_TYPE_F32, {ne0, ne1, 1, 1}));
+        }
+
+        exponent <<= 1;
+    }
 
     for (ggml_type type : {GGML_TYPE_F32, GGML_TYPE_F16}) {
         test_cases.emplace_back(new test_rope(type, {128,  32, 10, 1}, 128, 0, 512)); // llama 7B

tests/test-quantize-fns.cpp

@@ -134,6 +134,12 @@ int main(int argc, char * argv[]) {
             continue;
         }
 
+        const ggml_type ei = (ggml_type)i;
+        if (ei == GGML_TYPE_IQ2_XXS || ei == GGML_TYPE_IQ2_XS) {
+            printf("Skip %s due to missing quantization functionality\n", ggml_type_name(ei));
+            continue;
+        }
+
         printf("Testing %s\n", ggml_type_name((ggml_type) i));
 
         if (qfns.from_float && qfns.to_float) {