llama : remember and restore kv cache data pointers (#1104)

because their value is stored in buf and overwritten by memcpy

.github/workflows/build.yml

@@ -81,7 +81,6 @@ jobs:
       matrix:
         sanitizer: [ADDRESS, THREAD, UNDEFINED]
         build_type: [Debug, Release]
-        accelerate: [ON, OFF]
 
     steps:
       - name: Clone
@@ -99,7 +98,7 @@ jobs:
         run: |
           mkdir build
           cd build
-          cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }} -DLLAMA_ACCELERATE=${{ matrix.accelerate }}
+          cmake .. -DLLAMA_SANITIZE_${{ matrix.sanitizer }}=ON -DCMAKE_BUILD_TYPE=${{ matrix.build_type }}
           cmake --build . --config ${{ matrix.build_type }}
 
       - name: Test

Makefile

@@ -102,7 +102,7 @@ ifdef LLAMA_OPENBLAS
 endif
 ifdef LLAMA_CUBLAS
 	CFLAGS  += -DGGML_USE_CUBLAS -I/usr/local/cuda/include
-	LDFLAGS += -lcublas_static -lculibos -lcudart_static -lcublasLt_static -lpthread -ldl -lrt -L/usr/local/cuda/lib64
+	LDFLAGS += -lcublas -lculibos -lcudart -lcublasLt -lpthread -ldl -lrt -L/usr/local/cuda/lib64
 	OBJS	+= ggml-cuda.o
 ggml-cuda.o: ggml-cuda.cu ggml-cuda.h
 	nvcc -arch=native -c -o $@ $<

SHA256SUMS

@@ -1,12 +1,27 @@
 700df0d3013b703a806d2ae7f1bfb8e59814e3d06ae78be0c66368a50059f33d  models/7B/consolidated.00.pth
+666a4bb533b303bdaf89e1b6a3b6f93535d868de31d903afdc20983dc526c847  models/7B/ggml-model-f16.bin
+fcb7664c2e69776920b526362a243e912f73c36b1ec892eb354bab940f5edb5a  models/7B/ggml-model-q4_0.bin
+cc061458339a3eb8bcecbf0a825e9924fb7d1a8150f63cd5d091caa99215aafe  models/7B/ggml-model-q4_1.bin
+1bc7484c24a87612726d756f1761890e7acf5f412e23378577ce50fbe789b5b8  models/7B/ggml-model-q4_2.bin
+3429bf198ec771886cf81a574df45245f3ebf04f0ce0956b73ef5d0ab01ff48b  models/7B/ggml-model-q4_3.bin
 7e89e242ddc0dd6f060b43ca219ce8b3e8f08959a72cb3c0855df8bb04d46265  models/7B/params.json
 745bf4e29a4dd6f411e72976d92b452da1b49168a4f41c951cfcc8051823cf08  models/13B/consolidated.00.pth
 d5ccbcc465c71c0de439a5aeffebe8344c68a519bce70bc7f9f92654ee567085  models/13B/consolidated.01.pth
+2b206e9b21fb1076f11cafc624e2af97c9e48ea09312a0962153acc20d45f808  models/13B/ggml-model-f16.bin
+4b69e4d6b6e3275230955997b90407fceca7e5ab3daf2e63a2c9e7270a8e1e3e  models/13B/ggml-model-q4_0.bin
+d9581b5b88e5622532fe897c9f9b0e67a317d22dd27a6f90fa4ab8c6d23ccdbb  models/13B/ggml-model-q4_1.bin
+8d55a2077317ec9a928c7851d6a43e08e51f7e9e08360f2a7a7e1deefea3134f  models/13B/ggml-model-q4_2.bin
+4208cdec9788ffa48dc1a17af2c36a0299f5bf3eb0e2b87889dda7fad591fca3  models/13B/ggml-model-q4_3.bin
 4ab77bec4d4405ccb66a97b282574c89a94417e3c32e5f68f37e2876fc21322f  models/13B/params.json
 e23294a58552d8cdec5b7e8abb87993b97ea6eced4178ff2697c02472539d067  models/30B/consolidated.00.pth
 4e077b7136c7ae2302e954860cf64930458d3076fcde9443f4d0e939e95903ff  models/30B/consolidated.01.pth
 24a87f01028cbd3a12de551dcedb712346c0b5cbdeff1454e0ddf2df9b675378  models/30B/consolidated.02.pth
 1adfcef71420886119544949767f6a56cb6339b4d5fcde755d80fe68b49de93b  models/30B/consolidated.03.pth
+7e1b524061a9f4b27c22a12d6d2a5bf13b8ebbea73e99f218809351ed9cf7d37  models/30B/ggml-model-f16.bin
+7a679908ce31c9d6ae2e38d6059bcd4d0ad3a870cd58cc1c8f7b36f2b2f51c73  models/30B/ggml-model-q4_0.bin
+7b75ac615fa369ee593493a7e6ef87542bf0350255db928b22c5a24f6d598bcd  models/30B/ggml-model-q4_1.bin
+2c82b4954a94a6a284f452f6011c1e4f0d20362c194a0b1eb5737f5fd8a20fb3  models/30B/ggml-model-q4_2.bin
+a6188660199dbcb8d5658abe7d89169869e50423494385830d9e6b330ea7fc33  models/30B/ggml-model-q4_3.bin
 2c07118ea98d69dbe7810d88520e30288fa994751b337f8fca02b171955f44cb  models/30B/params.json
 135c563f6b3938114458183afb01adc9a63bef3d8ff7cccc3977e5d3664ecafe  models/65B/consolidated.00.pth
 9a600b37b19d38c7e43809485f70d17d1dc12206c07efa83bc72bb498a568bde  models/65B/consolidated.01.pth
@@ -16,5 +31,10 @@ e7babf7c5606f165a3756f527cb0fedc4f83e67ef1290391e52fb1cce5f26770  models/65B/con
 a287c0dfe49081626567c7fe87f74cce5831f58e459b427b5e05567641f47b78  models/65B/consolidated.05.pth
 72b4eba67a1a3b18cb67a85b70f8f1640caae9b40033ea943fb166bd80a7b36b  models/65B/consolidated.06.pth
 d27f5b0677d7ff129ceacd73fd461c4d06910ad7787cf217b249948c3f3bc638  models/65B/consolidated.07.pth
+60758f2384d74e423dffddfd020ffed9d3bb186ebc54506f9c4a787d0f5367b0  models/65B/ggml-model-f16.bin
+c671fe1bce71499ac732ec999770ebe53ac486623a7891e42c9dfdb6962d2c64  models/65B/ggml-model-q4_0.bin
+4743a28aac3e5f32a6e838a815f51d3779de44fbbe251d745251e66c23c5950f  models/65B/ggml-model-q4_1.bin
+4a145a210c56982389b1ed34387e0590c3e0d7325fa9be4f2284fe4d244a3633  models/65B/ggml-model-q4_2.bin
+305e91a4608b4f627b9b8ad5b4af75187d2684254bfd76dcb9db571618ef293c  models/65B/ggml-model-q4_3.bin
 999ed1659b469ccc2a941714c0a9656fa571d17c9f7c8c7589817ca90edef51b  models/65B/params.json
 9e556afd44213b6bd1be2b850ebbbd98f5481437a8021afaf58ee7fb1818d347  models/tokenizer.model

examples/perplexity/perplexity.cpp

@@ -53,7 +53,13 @@ void perplexity(llama_context * ctx, const gpt_params & params) {
         auto end_t = std::chrono::high_resolution_clock::now();
         if (i == 0) {
             const float seconds = std::chrono::duration<float>(end_t - start_t).count();
-            printf("%.2f seconds per pass - ETA %.2f hours\n", seconds, (seconds * seq_count) / (60.0*60.0));
+            printf("%.2f seconds per pass - ETA ", seconds);
+            int total_seconds = (int)(seconds * seq_count);
+            if (total_seconds >= 60*60) {
+                printf("%d hours ", total_seconds / (60*60));
+                total_seconds = total_seconds % (60*60);
+            }
+            printf("%d minutes\n", total_seconds / 60);
         }
         // We get the logits for all the tokens in the context window (params.n_ctx)
         // from llama_eval above.  Now, based on https://huggingface.co/docs/transformers/perplexity,

examples/quantize-stats/quantize-stats.cpp

@@ -15,6 +15,8 @@
 #include <string>
 #include <unordered_map>
 #include <vector>
+#include <thread>
+#include <mutex>
 
 struct quantize_stats_params {
     std::string model = "models/7B/ggml-model-f16.bin";
@@ -27,7 +29,6 @@ struct quantize_stats_params {
     std::vector<enum ggml_type> include_types;
 };
 
-const int64_t SCRATCH_ELEMENTS = 32*32;
 const size_t HISTOGRAM_BUCKETS = 150;
 const double HISTOGRAM_RANGE = 0.03;
 
@@ -90,6 +91,13 @@ void update_error_stats(int64_t nelements, const float * input, const float * ou
     stats.num_samples += nelements;
 }
 
+void combine_error_stats(error_stats & into, const error_stats & from) {
+    into.num_samples += from.num_samples;
+    into.total_error += from.total_error;
+    if (from.max_error > into.max_error) into.max_error = from.max_error;
+    for (size_t i=0; i<HISTOGRAM_BUCKETS; ++i) into.error_histogram[i] += from.error_histogram[i];
+}
+
 double find_quantile(const error_stats & stats, double quantile) {
     double sum = std::accumulate(std::begin(stats.error_histogram), std::end(stats.error_histogram), 0.0);
 
@@ -130,6 +138,36 @@ static bool tensor_is_contiguous(const struct ggml_tensor * tensor) {
         tensor->nb[3] == tensor->nb[2]*tensor->ne[2];
 }
 
+void test_roundtrip_on_chunk(
+        const ggml_tensor * layer,
+        int64_t offset,
+        int64_t chunk_size,
+        const quantize_fns_t & qfns,
+        bool use_reference,
+        float * input_scratch,
+        char * quantized_scratch,
+        float * output_scratch,
+        error_stats & stats) {
+
+    if (layer->type == GGML_TYPE_F16) {
+        for (int i = 0; i < chunk_size; i++) {
+            input_scratch[i] = ggml_get_f32_1d(layer, i + offset);
+        }
+    } else {
+        input_scratch = ggml_get_data_f32(layer) + offset;
+    }
+
+    if (use_reference) {
+        qfns.quantize_row_q_reference(input_scratch, quantized_scratch, chunk_size);
+    } else {
+        qfns.quantize_row_q(input_scratch, quantized_scratch, chunk_size);
+    }
+    qfns.dequantize_row_q(quantized_scratch, output_scratch, chunk_size);
+
+    update_error_stats(chunk_size, input_scratch, output_scratch, stats);
+}
+
+
 // Run quantization function for a single layer and update error stats
 void test_roundtrip_on_layer(
         std::string & name,
@@ -137,40 +175,61 @@ void test_roundtrip_on_layer(
         const quantize_fns_t & qfns,
         bool use_reference,
         const ggml_tensor * layer,
-        float * input_scratch,
-        char *quantized_scratch,
-        float * output_scratch,
-        error_stats & total_error) {
+        std::vector<float> & input_scratch,
+        std::vector<char> & quantized_scratch,
+        std::vector<float> & output_scratch,
+        error_stats & total_error,
+        int max_thread = 0) {
 
     assert(tensor_is_contiguous(layer));
     error_stats layer_error {};
-    int64_t nelements = ggml_nelements(layer);
+    uint64_t nelements = ggml_nelements(layer);
 
-    for (int64_t offset = 0; offset < nelements; offset += SCRATCH_ELEMENTS) {
-        int64_t chunk_size = std::min(SCRATCH_ELEMENTS, nelements - offset);
-
-        if (layer->type == GGML_TYPE_F16) {
-            for (int i = 0; i < chunk_size; i++) {
-                input_scratch[i] = ggml_get_f32_1d(layer, i + offset);
-            }
-        } else {
-            input_scratch = ggml_get_data_f32(layer) + offset;
-        }
-
-        if (use_reference) {
-            qfns.quantize_row_q_reference(input_scratch, quantized_scratch, chunk_size);
-        } else {
-            qfns.quantize_row_q(input_scratch, quantized_scratch, chunk_size);
-        }
-        qfns.dequantize_row_q(quantized_scratch, output_scratch, chunk_size);
-
-        update_error_stats(chunk_size, input_scratch, output_scratch, total_error);
-        if (print_layer_stats) {
-            update_error_stats(chunk_size, input_scratch, output_scratch, layer_error);
-        }
+    float* input_scratch_ptr = nullptr;
+    if (layer->type == GGML_TYPE_F16) {
+        if (input_scratch.size() < nelements) input_scratch.resize(nelements);
+        input_scratch_ptr = input_scratch.data();
     }
+    if (quantized_scratch.size() < 4*nelements) quantized_scratch.resize(4*nelements);
+    if (output_scratch.size() < nelements) output_scratch.resize(nelements);
+
+    if (max_thread < 1) max_thread = std::thread::hardware_concurrency();
+    int chunk_size = 32*512;
+    int num_chunks = (nelements + chunk_size - 1)/chunk_size;
+
+    if (num_chunks < 2 || max_thread < 2) {
+        test_roundtrip_on_chunk(layer, 0, nelements, qfns, use_reference, input_scratch_ptr, quantized_scratch.data(),
+                output_scratch.data(), print_layer_stats ? layer_error : total_error);
+    } else {
+        auto & stats = print_layer_stats ? layer_error : total_error;
+        std::mutex mutex;
+        uint64_t counter = 0;
+        auto compute = [&mutex, &counter, &stats, &qfns, nelements, layer, use_reference, input_scratch_ptr,
+             &quantized_scratch, &output_scratch, chunk_size] () {
+            error_stats local_stats {};
+            while (true) {
+                std::unique_lock<std::mutex> lock(mutex);
+                uint64_t offset = counter; counter += chunk_size;
+                if (offset >= nelements) {
+                    combine_error_stats(stats, local_stats);
+                    break;
+                }
+                lock.unlock();
+                uint64_t chunk = offset + chunk_size < nelements ? chunk_size : nelements - offset;
+                test_roundtrip_on_chunk(layer, offset, chunk, qfns, use_reference, input_scratch_ptr + offset,
+                        quantized_scratch.data() + 4*offset, output_scratch.data() + offset, local_stats);
+            }
+        };
+        int nthread = std::min(num_chunks, max_thread);
+        std::vector<std::thread> workers(nthread-1);
+        for (auto& w : workers) w = std::thread(compute);
+        compute();
+        for (auto& w : workers) w.join();
+    }
+
     if (print_layer_stats) {
         print_error_stats(name, layer_error, false);
+        combine_error_stats(total_error, layer_error);
     }
 }
 
@@ -181,6 +240,7 @@ int main(int argc, char ** argv) {
 
     // read command line
 
+    int max_thread = 0;
     bool invalid_param = false;
     std::string arg;
     for (int i = 1; i < argc; i++) {
@@ -230,6 +290,12 @@ int main(int argc, char ** argv) {
                 fprintf(stderr, "error: %s not in list of types\n", argv[i]);
                 invalid_param = true;
             }
+        } else if (arg == "-n" || arg == "--num-threads") {
+            if (++i >= argc) {
+                invalid_param = true;
+                break;
+            }
+            max_thread = atoi(argv[i]);
         } else {
             fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
             quantize_stats_print_usage(argc, argv);
@@ -295,9 +361,9 @@ int main(int argc, char ** argv) {
     }
     printf("testing %d layers with max size %" PRId64 "\n", included_layers, max_nelements);
     // allocate scratch space
-    std::vector<float> input_scratch(SCRATCH_ELEMENTS);
-    std::vector<char> quantized_scratch(SCRATCH_ELEMENTS*4);
-    std::vector<float> output_scratch(SCRATCH_ELEMENTS);
+    std::vector<float> input_scratch;
+    std::vector<char> quantized_scratch;
+    std::vector<float> output_scratch;
 
     // loop throught quantization types
     for (int i = 0; i < GGML_TYPE_COUNT; i++) {
@@ -328,10 +394,11 @@ int main(int argc, char ** argv) {
                         qfns,
                         params.reference,
                         kv_tensor.second,
-                        input_scratch.data(),
-                        quantized_scratch.data(),
-                        output_scratch.data(),
-                        global_stats
+                        input_scratch,
+                        quantized_scratch,
+                        output_scratch,
+                        global_stats,
+                        max_thread
                 );
             }
 

examples/quantize/quantize.cpp

@@ -10,11 +10,12 @@
 int main(int argc, char ** argv) {
     ggml_time_init();
 
-    if (argc != 4) {
-        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+    if (argc < 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type [nthread]\n", argv[0]);
         fprintf(stderr, "  type = %d - q4_0\n", LLAMA_FTYPE_MOSTLY_Q4_0);
         fprintf(stderr, "  type = %d - q4_1\n", LLAMA_FTYPE_MOSTLY_Q4_1);
         fprintf(stderr, "  type = %d - q4_2\n", LLAMA_FTYPE_MOSTLY_Q4_2);
+        fprintf(stderr, "  type = %d - q4_3\n", LLAMA_FTYPE_MOSTLY_Q4_3);
         return 1;
     }
 
@@ -29,6 +30,7 @@ int main(int argc, char ** argv) {
     const std::string fname_out = argv[2];
 
     const enum llama_ftype ftype = (enum llama_ftype)atoi(argv[3]);
+    int nthread = argc > 4 ? atoi(argv[4]) : 0;
 
     const int64_t t_main_start_us = ggml_time_us();
 
@@ -38,7 +40,7 @@ int main(int argc, char ** argv) {
     {
         const int64_t t_start_us = ggml_time_us();
 
-        if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), ftype)) {
+        if (llama_model_quantize(fname_inp.c_str(), fname_out.c_str(), ftype, nthread)) {
             fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
             return 1;
         }

ggml-cuda.cu

@@ -22,11 +22,20 @@ static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 b
 
 #define QK4_2 16
 typedef struct {
-    __half d;               // delta
+    __half  d;              // delta
     uint8_t qs[QK4_2 / 2];  // nibbles / quants
 } block_q4_2;
 static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding");
 
+#define QK4_3 16
+typedef struct {
+    __half  d;         // delta
+    __half  m;         // min
+    uint8_t qs[QK4_3 / 2]; // nibbles / quants
+} block_q4_3;
+static_assert(sizeof(block_q4_3) == 2 * sizeof(ggml_fp16_t) + QK4_3 / 2, "wrong q4_3 block size/padding");
+
+
 
 static __global__ void dequantize_block_q4_0(const void * vx, float * y) {
     const block_q4_0 * x = (const block_q4_0 *) vx;
@@ -98,6 +107,30 @@ static __global__ void dequantize_block_q4_2(const void * vx, float * y) {
     }
 }
 
+static __global__ void dequantize_block_q4_3(const void * vx, float * y) {
+    const block_q4_3 * x = (const block_q4_3 *) vx;
+
+    const int i = blockIdx.x;
+
+    const float d = x[i].d;
+    const float m = x[i].m;
+
+    const uint8_t * pp = x[i].qs;
+
+    for (int l = 0; l < QK4_3; l += 2) {
+        const uint8_t vi = pp[l/2];
+
+        const int8_t vi0 = vi & 0xf;
+        const int8_t vi1 = vi >> 4;
+
+        const float v0 = vi0*d + m;
+        const float v1 = vi1*d + m;
+
+        y[i*QK4_3 + l + 0] = v0;
+        y[i*QK4_3 + l + 1] = v1;
+    }
+}
+
 extern "C" {
     __host__ void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
         const int nb = k / QK4_0;
@@ -113,4 +146,9 @@ extern "C" {
         const int nb = k / QK4_2;
         dequantize_block_q4_2<<<nb, 1, 0, stream>>>(vx, y);
     }
+
+    __host__ void dequantize_row_q4_3_cuda(const void * vx, float * y, int k, cudaStream_t stream) {
+        const int nb = k / QK4_3;
+        dequantize_block_q4_3<<<nb, 1, 0, stream>>>(vx, y);
+    }
 }

ggml-cuda.h

@@ -5,6 +5,7 @@ extern "C" {
 void dequantize_row_q4_0_cuda(const void * vx, float * y, int k, cudaStream_t stream);
 void dequantize_row_q4_1_cuda(const void * vx, float * y, int k, cudaStream_t stream);
 void dequantize_row_q4_2_cuda(const void * vx, float * y, int k, cudaStream_t stream);
+void dequantize_row_q4_3_cuda(const void * vx, float * y, int k, cudaStream_t stream);
 
 #ifdef  __cplusplus
 }

ggml.c

@@ -637,7 +637,7 @@ typedef struct {
     float   m;          // min
     uint8_t qs[QK4_1 / 2];  // nibbles / quants
 } block_q4_1;
-static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
+static_assert(sizeof(block_q4_1) == 2 * sizeof(float) + QK4_1 / 2, "wrong q4_1 block size/padding");
 
 #define QK4_2 16
 typedef struct {
@@ -646,12 +646,21 @@ typedef struct {
 } block_q4_2;
 static_assert(sizeof(block_q4_2) == sizeof(ggml_fp16_t) + QK4_2 / 2, "wrong q4_2 block size/padding");
 
+#define QK4_3 16
+typedef struct {
+    ggml_fp16_t d;         // delta
+    ggml_fp16_t m;         // min
+    uint8_t qs[QK4_3 / 2]; // nibbles / quants
+} block_q4_3;
+static_assert(sizeof(block_q4_3) == 2 * sizeof(ggml_fp16_t) + QK4_3 / 2, "wrong q4_3 block size/padding");
+
 #define QK8_0 32
 typedef struct {
     float   d;          // delta
+    float   s;          // d * sum(qs[i])
     int8_t  qs[QK8_0];  // quants
 } block_q8_0;
-static_assert(sizeof(block_q8_0) == sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
+static_assert(sizeof(block_q8_0) == 2*sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
 
 
 // reference implementation for deterministic creation of model files
@@ -1203,7 +1212,6 @@ static void quantize_row_q4_2_rmse(const float * restrict x, block_q4_2 * restri
     const int nb = k / QK4_2;
 
     for (int i = 0; i < nb; i++) {
-
         float scale = kquantize_q4_with_bounds(QK4_2, -8, 7, x, CANDIDATE_COUNT, candidates, L);
         y[i].d = GGML_FP32_TO_FP16(scale);
 
@@ -1231,6 +1239,49 @@ static void quantize_row_q4_2(const float * restrict x, void * restrict vy, int
     quantize_row_q4_2_rmse(x, y, k);
 }
 
+static void quantize_row_q4_3_reference(const float * restrict x, block_q4_3 * restrict y, int k) {
+    assert(k % QK4_3 == 0);
+    const int nb = k / QK4_3;
+
+    for (int i = 0; i < nb; i++) {
+        float min = FLT_MAX;
+        float max = -FLT_MAX;
+
+        for (int l = 0; l < QK4_3; l++) {
+            const float v = x[i*QK4_3 + l];
+            if (v < min) min = v;
+            if (v > max) max = v;
+        }
+
+        const float d = (max - min) / ((1 << 4) - 1);
+        const float id = d ? 1.0f/d : 0.0f;
+
+        y[i].d = GGML_FP32_TO_FP16(d);
+        y[i].m = GGML_FP32_TO_FP16(min);
+
+        for (int l = 0; l < QK4_3; l += 2) {
+            const float v0 = (x[i*QK4_3 + l + 0] - min)*id;
+            const float v1 = (x[i*QK4_3 + l + 1] - min)*id;
+
+            const uint8_t vi0 = (int) (v0 + 0.5f);
+            const uint8_t vi1 = (int) (v1 + 0.5f);
+
+            assert(vi0 < 16);
+            assert(vi1 < 16);
+
+            y[i].qs[l/2] = vi0 | (vi1 << 4);
+        }
+    }
+}
+
+static void quantize_row_q4_3(const float * restrict x, void * restrict vy, int k) {
+    assert(k % QK4_3 == 0);
+
+    block_q4_3 * restrict y = vy;
+
+    quantize_row_q4_3_reference(x, y, k);
+}
+
 // reference implementation for deterministic creation of model files
 static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * restrict y, int k) {
     assert(k % QK8_0 == 0);
@@ -1249,13 +1300,39 @@ static void quantize_row_q8_0_reference(const float * restrict x, block_q8_0 * r
 
         y[i].d = d;
 
+        int sum = 0;
         for (int l = 0; l < QK8_0; ++l) {
             const float v = x[i*QK8_0 + l]*id;
             y[i].qs[l] = roundf(v);
+            sum += y[i].qs[l];
         }
+        y[i].s = d * sum;
     }
 }
 
+#ifdef __AVX2__
+// There is no better way of doing this?
+// I guess not, AVX is not very good at horizontal sums.
+// The commented solution for a hotrizontal sum was suggested by @pubby as being slightly
+// faster than the solution below. As I don't have an AVX2 system handt right now to test,
+// keeping the original.
+// TODO: Please try and if it does make a differece, uncomment and remove the implementation below.
+//static inline float horizontal_sum(__m256i a) {
+//    __m256i b = _mm256_castps_si256(_mm256_movehdup_ps(_mm256_castsi256_ps(a)));
+//    __m256i sum = _mm256_add_epi32(a, b);
+//    __m256i hi = _mm256_unpackhi_epi64(sum, sum);
+//    sum = _mm256_add_epi32(sum, hi);
+//    return _mm256_cvtsi256_si32(sum) + _mm256_extract_epi32(sum, 4);
+//}
+static inline float horizontal_sum(__m256i a) {
+    __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extracti128_si256(a, 1));
+    __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128);
+    __m128i sum64 = _mm_add_epi32(hi64, sum128);
+    __m128i hi32  = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
+    return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
+}
+#endif
+
 static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int k) {
     assert(k % QK8_0 == 0);
     const int nb = k / QK8_0;
@@ -1282,6 +1359,8 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
 
         y[i].d = d;
 
+        int32x4_t accv = vdupq_n_s32(0);
+
         for (int l = 0; l < 8; l++) {
             const float32x4_t v  = vmulq_n_f32(srcv[l], id);
             const int32x4_t   vi = vcvtnq_s32_f32(v);
@@ -1290,7 +1369,11 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
             y[i].qs[4*l + 1] = vgetq_lane_s32(vi, 1);
             y[i].qs[4*l + 2] = vgetq_lane_s32(vi, 2);
             y[i].qs[4*l + 3] = vgetq_lane_s32(vi, 3);
+
+            accv = vaddq_s32(accv, vi);
         }
+        int32_t sum = vaddvq_s32(accv);
+        y[i].s = d * sum;
     }
 #elif defined(__AVX2__) || defined(__AVX__)
     for (int i = 0; i < nb; i++) {
@@ -1338,6 +1421,10 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
         __m256i i3 = _mm256_cvtps_epi32( v3 );
 
 #if defined(__AVX2__)
+
+        // Compute the sum of the quants and set y[i].s
+        y[i].s = d * horizontal_sum(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)));
+
         // Convert int32 to int16
         i0 = _mm256_packs_epi32( i0, i1 );	// 0, 1, 2, 3,  8, 9, 10, 11,  4, 5, 6, 7, 12, 13, 14, 15
         i2 = _mm256_packs_epi32( i2, i3 );	// 16, 17, 18, 19,  24, 25, 26, 27,  20, 21, 22, 23, 28, 29, 30, 31
@@ -1380,6 +1467,14 @@ static void quantize_row_q8_0(const float * restrict x, void * restrict vy, int
     // scalar
     quantize_row_q8_0_reference(x, y, k);
 #endif
+#if defined __AVX__
+    // TODO: vectorize this
+    for (int i=0; i<nb; ++i) {
+        int sum = 0;
+        for (int l=0; l<QK8_0; ++l) sum += y[i].qs[l];
+        y[i].s = y[i].d * sum;
+    }
+#endif
 }
 
 static void dequantize_row_q4_0(const void * restrict vx, float * restrict y, int k) {
@@ -1635,9 +1730,40 @@ static void dequantize_row_q4_2(const void * restrict vx, float * restrict y, in
     }
 }
 
+static void dequantize_row_q4_3(const void * restrict vx, float * restrict y, int k) {
+    assert(k % QK4_3 == 0);
+    const int nb = k / QK4_3;
+
+    const block_q4_3 * restrict x = vx;
+
+    for (int i = 0; i < nb; i++) {
+        const float d = GGML_FP16_TO_FP32(x[i].d);
+        const float m = GGML_FP16_TO_FP32(x[i].m);
+
+        const uint8_t * restrict pp = x[i].qs;
+
+        for (int l = 0; l < QK4_3; l += 2) {
+            const uint8_t vi = pp[l/2];
+
+            const int8_t vi0 = vi & 0xf;
+            const int8_t vi1 = vi >> 4;
+
+            const float v0 = vi0*d + m;
+            const float v1 = vi1*d + m;
+
+            y[i*QK4_3 + l + 0] = v0;
+            y[i*QK4_3 + l + 1] = v1;
+
+            assert(!isnan(y[i*QK4_3 + l + 0]));
+            assert(!isnan(y[i*QK4_3 + l + 1]));
+        }
+    }
+}
+
 static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy);
+static void ggml_vec_dot_q4_3_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy);
 
 static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = {
     [GGML_TYPE_Q4_0] = {
@@ -1661,6 +1787,13 @@ static const quantize_fns_t quantize_fns[GGML_TYPE_COUNT] = {
         .quantize_row_q_dot       = quantize_row_q8_0,
         .vec_dot_q                = ggml_vec_dot_q4_2_q8_0,
     },
+    [GGML_TYPE_Q4_3] = {
+        .dequantize_row_q         = dequantize_row_q4_3,
+        .quantize_row_q           = quantize_row_q4_3,
+        .quantize_row_q_reference = (quantize_row_q_t) quantize_row_q4_3_reference, // TODO: RMSE optimization
+        .quantize_row_q_dot       = quantize_row_q8_0,
+        .vec_dot_q                = ggml_vec_dot_q4_3_q8_0,
+    },
     [GGML_TYPE_Q8_0] = {
         .dequantize_row_q         = NULL,   // TODO
         .quantize_row_q           = quantize_row_q8_0,
@@ -2284,14 +2417,17 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
+    float sum8 = 0;
+
     for (int i = 0; i < nb; i += 2) {
         const block_q4_0 * restrict x0 = &x[i + 0];
         const block_q4_0 * restrict x1 = &x[i + 1];
         const block_q8_0 * restrict y0 = &y[i + 0];
         const block_q8_0 * restrict y1 = &y[i + 1];
 
+        sum8 += x0->d * y0->s + x1->d * y1->s;
+
         const uint8x16_t m4b   = vdupq_n_u8(0xf);
-        const int8x16_t  s8b   = vdupq_n_s8(0x8);
 
         const uint8x16_t v0_0 = vld1q_u8(x0->qs);
         const uint8x16_t v0_1 = vld1q_u8(x1->qs);
@@ -2302,12 +2438,6 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
         const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
         const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
 
-        // sub 8
-        const int8x16_t v0_0ls = vsubq_s8(v0_0l, s8b);
-        const int8x16_t v0_0hs = vsubq_s8(v0_0h, s8b);
-        const int8x16_t v0_1ls = vsubq_s8(v0_1l, s8b);
-        const int8x16_t v0_1hs = vsubq_s8(v0_1h, s8b);
-
         // load y
         const int8x16_t v1_0l = vld1q_s8(y0->qs);
         const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
@@ -2322,21 +2452,21 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
 
 #if defined(__ARM_FEATURE_DOTPROD)
         // dot product into int32x4_t
-        const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0ls, v1_0ls), v0_0hs, v1_0hs);
-        const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1ls, v1_1ls), v0_1hs, v1_1hs);
+        const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0ls), v0_0h, v1_0hs);
+        const int32x4_t p_1 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1ls), v0_1h, v1_1hs);
 
         sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), x0->d*y0->d);
         sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), x1->d*y1->d);
 #else
-        const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0ls), vget_low_s8 (v1_0ls));
-        const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0ls), vget_high_s8(v1_0ls));
-        const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hs), vget_low_s8 (v1_0hs));
-        const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hs), vget_high_s8(v1_0hs));
+        const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0l), vget_low_s8 (v1_0ls));
+        const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0l), vget_high_s8(v1_0ls));
+        const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0h), vget_low_s8 (v1_0hs));
+        const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0h), vget_high_s8(v1_0hs));
 
-        const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1ls), vget_low_s8 (v1_1ls));
-        const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1ls), vget_high_s8(v1_1ls));
-        const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hs), vget_low_s8 (v1_1hs));
-        const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hs), vget_high_s8(v1_1hs));
+        const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1l), vget_low_s8 (v1_1ls));
+        const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1l), vget_high_s8(v1_1ls));
+        const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1h), vget_low_s8 (v1_1hs));
+        const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1h), vget_high_s8(v1_1hs));
 
         const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
         const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
@@ -2348,7 +2478,7 @@ static void ggml_vec_dot_q4_0_q8_0(const int n, float * restrict s, const void *
 #endif
     }
 
-    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) - 8 * sum8;
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
@@ -2481,12 +2611,16 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
     float32x4_t sumv0 = vdupq_n_f32(0.0f);
     float32x4_t sumv1 = vdupq_n_f32(0.0f);
 
+    float summs = 0;
+
     for (int i = 0; i < nb; i += 2) {
         const block_q4_1 * restrict x0 = &x[i + 0];
         const block_q4_1 * restrict x1 = &x[i + 1];
         const block_q8_0 * restrict y0 = &y[i + 0];
         const block_q8_0 * restrict y1 = &y[i + 1];
 
+        summs += x0->m * y0->s + x1->m * y1->s;
+
         const uint8x16_t m4b = vdupq_n_u8(0xf);
 
         const uint8x16_t v0_0 = vld1q_u8(x0->qs);
@@ -2510,17 +2644,6 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
         const int8x16_t v1_1ls = vuzp1q_s8(v1_1l, v1_1h);
         const int8x16_t v1_1hs = vuzp2q_s8(v1_1l, v1_1h);
 
-        const int16x8_t s0i = vaddq_s16(
-                        vaddq_s16(vmovl_s8(vget_low_s8(v1_0ls)), vmovl_s8(vget_high_s8(v1_0ls))),
-                        vaddq_s16(vmovl_s8(vget_low_s8(v1_0hs)), vmovl_s8(vget_high_s8(v1_0hs))));
-
-        const int16x8_t s1i = vaddq_s16(
-                        vaddq_s16(vmovl_s8(vget_low_s8(v1_1ls)), vmovl_s8(vget_high_s8(v1_1ls))),
-                        vaddq_s16(vmovl_s8(vget_low_s8(v1_1hs)), vmovl_s8(vget_high_s8(v1_1hs))));
-
-        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddl_s16(vget_low_s16(s0i), vget_high_s16(s0i))), x0->m*y0->d);
-        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddl_s16(vget_low_s16(s1i), vget_high_s16(s1i))), x1->m*y1->d);
-
 #if defined(__ARM_FEATURE_DOTPROD)
         // dot product into int32x4_t
         const int32x4_t p_0 = vdotq_s32(vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0ls), v0_0h, v1_0hs);
@@ -2549,24 +2672,26 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
 #endif
     }
 
-    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
 #elif defined(__AVX2__)
     // Initialize accumulator with zeros
     __m256 acc = _mm256_setzero_ps();
 
+    float summs = 0;
+
     // Main loop
     for (int i = 0; i < nb; ++i) {
         const float * d0 = &x[i].d;
         const float * d1 = &y[i].d;
-        const float * m0 = &x[i].m;
+        //const float * m0 = &x[i].m;
+
+        summs += x[i].m * y[i].s;
 
         const __m256 d0v = _mm256_broadcast_ss( d0 );
         const __m256 d1v = _mm256_broadcast_ss( d1 );
-        const __m256 m0v = _mm256_broadcast_ss( m0 );
 
         // Compute combined scales
         const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
-        const __m256 d1m0 = _mm256_mul_ps( d1v, m0v );
 
         // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
         const __m256i bx = bytes_from_nibbles_32(x[i].qs);
@@ -2588,15 +2713,6 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
 
         // Accumulate d0*d1*x*y
         acc = _mm256_fmadd_ps( d0d1, xy, acc );
-
-        // Compute sum of y values
-        const __m256i y16_l = _mm256_cvtepi8_epi16( _mm256_castsi256_si128( by ) );
-        const __m256i y16_h = _mm256_cvtepi8_epi16( _mm256_extracti128_si256( by, 1 ) );
-        const __m256i ysumi = _mm256_madd_epi16( _mm256_add_epi16(y16_l, y16_h), ones );
-        const __m256 ysum = _mm256_cvtepi32_ps( ysumi );
-
-        // Accumulate d1*m0*y
-        acc = _mm256_fmadd_ps( d1m0, ysum, acc );
     }
 
     // Return horizontal sum of the acc vector
@@ -2605,7 +2721,7 @@ static void ggml_vec_dot_q4_1_q8_0(const int n, float * restrict s, const void *
     res = _mm_add_ps( res, _mm_movehl_ps( res, res ) );
     res = _mm_add_ss( res, _mm_movehdup_ps( res ) );
 
-    sumf = _mm_cvtss_f32( res );
+    sumf = _mm_cvtss_f32( res ) + summs;
 #else
     // scalar
     for (int i = 0; i < nb; i++) {
@@ -2655,6 +2771,7 @@ static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void *
         const block_q4_2 * restrict x0_1 = &x[2*(i + 0) + 1];
         const block_q4_2 * restrict x1_0 = &x[2*(i + 1) + 0];
         const block_q4_2 * restrict x1_1 = &x[2*(i + 1) + 1];
+
         const block_q8_0 * restrict y0 = &y[i + 0];
         const block_q8_0 * restrict y1 = &y[i + 1];
 
@@ -2809,6 +2926,154 @@ static void ggml_vec_dot_q4_2_q8_0(const int n, float * restrict s, const void *
     *s = sumf;
 }
 
+static void ggml_vec_dot_q4_3_q8_0(const int n, float * restrict s, const void * restrict vx, const void * restrict vy) {
+    const int nb = n / QK8_0;
+
+    assert(n % QK8_0 == 0);
+    assert(nb % 2 == 0);
+    assert(QK8_0 == 2*QK4_2);
+
+    const block_q4_3 * restrict x = vx;
+    const block_q8_0 * restrict y = vy;
+
+    float sumf = 0.0;
+
+#if defined(__ARM_NEON)
+    float32x4_t sumv0 = vdupq_n_f32(0.0f);
+    float32x4_t sumv1 = vdupq_n_f32(0.0f);
+
+    for (int i = 0; i < nb; i += 2) {
+        const block_q4_3 * restrict x0_0 = &x[2*(i + 0) + 0];
+        const block_q4_3 * restrict x0_1 = &x[2*(i + 0) + 1];
+        const block_q4_3 * restrict x1_0 = &x[2*(i + 1) + 0];
+        const block_q4_3 * restrict x1_1 = &x[2*(i + 1) + 1];
+
+        const block_q8_0 * restrict y0 = &y[i + 0];
+        const block_q8_0 * restrict y1 = &y[i + 1];
+
+        const uint8x16_t m4b = vdupq_n_u8(0xf);
+
+        const float x0_0d = GGML_FP16_TO_FP32(x0_0->d);
+        const float x0_1d = GGML_FP16_TO_FP32(x0_1->d);
+        const float x1_0d = GGML_FP16_TO_FP32(x1_0->d);
+        const float x1_1d = GGML_FP16_TO_FP32(x1_1->d);
+
+        const float x0_0m = GGML_FP16_TO_FP32(x0_0->m);
+        const float x0_1m = GGML_FP16_TO_FP32(x0_1->m);
+        const float x1_0m = GGML_FP16_TO_FP32(x1_0->m);
+        const float x1_1m = GGML_FP16_TO_FP32(x1_1->m);
+
+        const uint8x16_t v0_0 = vcombine_u8(vld1_u8(x0_0->qs), vld1_u8(x0_1->qs));
+        const uint8x16_t v0_1 = vcombine_u8(vld1_u8(x1_0->qs), vld1_u8(x1_1->qs));
+
+        // 4-bit -> 8-bit
+        const int8x16_t v0_0l = vreinterpretq_s8_u8(vandq_u8  (v0_0, m4b));
+        const int8x16_t v0_0h = vreinterpretq_s8_u8(vshrq_n_u8(v0_0, 4));
+        const int8x16_t v0_1l = vreinterpretq_s8_u8(vandq_u8  (v0_1, m4b));
+        const int8x16_t v0_1h = vreinterpretq_s8_u8(vshrq_n_u8(v0_1, 4));
+
+        // interleave
+        const int8x16_t v0_0lz = vzip1q_s8(v0_0l, v0_0h);
+        const int8x16_t v0_0hz = vzip2q_s8(v0_0l, v0_0h);
+        const int8x16_t v0_1lz = vzip1q_s8(v0_1l, v0_1h);
+        const int8x16_t v0_1hz = vzip2q_s8(v0_1l, v0_1h);
+
+        // load y
+        const int8x16_t v1_0l = vld1q_s8(y0->qs);
+        const int8x16_t v1_0h = vld1q_s8(y0->qs + 16);
+        const int8x16_t v1_1l = vld1q_s8(y1->qs);
+        const int8x16_t v1_1h = vld1q_s8(y1->qs + 16);
+
+        const int16x8_t sy0_0 = vaddq_s16(vmovl_s8(vget_low_s8(v1_0l)), vmovl_s8(vget_high_s8(v1_0l)));
+        const int16x8_t sy0_1 = vaddq_s16(vmovl_s8(vget_low_s8(v1_0h)), vmovl_s8(vget_high_s8(v1_0h)));
+
+        const int16x8_t sy1_0 = vaddq_s16(vmovl_s8(vget_low_s8(v1_1l)), vmovl_s8(vget_high_s8(v1_1l)));
+        const int16x8_t sy1_1 = vaddq_s16(vmovl_s8(vget_low_s8(v1_1h)), vmovl_s8(vget_high_s8(v1_1h)));
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy0_0), vget_high_s16(sy0_0))), x0_0m*y0->d);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy0_1), vget_high_s16(sy0_1))), x0_1m*y0->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy1_0), vget_high_s16(sy1_0))), x1_0m*y1->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddl_s16(vget_low_s16(sy1_1), vget_high_s16(sy1_1))), x1_1m*y1->d);
+
+#if defined(__ARM_FEATURE_DOTPROD)
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0lz, v1_0l)), x0_0d*y0->d);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_0hz, v1_0h)), x0_1d*y0->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1lz, v1_1l)), x1_0d*y1->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vdotq_s32(vdupq_n_s32(0), v0_1hz, v1_1h)), x1_1d*y1->d);
+#else
+        const int16x8_t pl0l = vmull_s8(vget_low_s8 (v0_0lz), vget_low_s8 (v1_0l));
+        const int16x8_t pl0h = vmull_s8(vget_high_s8(v0_0lz), vget_high_s8(v1_0l));
+        const int16x8_t ph0l = vmull_s8(vget_low_s8 (v0_0hz), vget_low_s8 (v1_0h));
+        const int16x8_t ph0h = vmull_s8(vget_high_s8(v0_0hz), vget_high_s8(v1_0h));
+
+        const int16x8_t pl1l = vmull_s8(vget_low_s8 (v0_1lz), vget_low_s8 (v1_1l));
+        const int16x8_t pl1h = vmull_s8(vget_high_s8(v0_1lz), vget_high_s8(v1_1l));
+        const int16x8_t ph1l = vmull_s8(vget_low_s8 (v0_1hz), vget_low_s8 (v1_1h));
+        const int16x8_t ph1h = vmull_s8(vget_high_s8(v0_1hz), vget_high_s8(v1_1h));
+
+        const int32x4_t pl0 = vaddq_s32(vpaddlq_s16(pl0l), vpaddlq_s16(pl0h));
+        const int32x4_t ph0 = vaddq_s32(vpaddlq_s16(ph0l), vpaddlq_s16(ph0h));
+        const int32x4_t pl1 = vaddq_s32(vpaddlq_s16(pl1l), vpaddlq_s16(pl1h));
+        const int32x4_t ph1 = vaddq_s32(vpaddlq_s16(ph1l), vpaddlq_s16(ph1h));
+
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(pl0), x0_0d*y0->d);
+        sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(ph0), x0_1d*y0->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(pl1), x1_0d*y1->d);
+        sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(ph1), x1_1d*y1->d);
+#endif
+    }
+
+    sumf = vaddvq_f32(sumv0) + vaddvq_f32(sumv1);
+#else
+    // scalar
+    for (int i = 0; i < nb; i++) {
+        const uint8_t * restrict x0 = x[2*i + 0].qs;
+        const uint8_t * restrict x1 = x[2*i + 1].qs;
+        const  int8_t * restrict y0 = y[i].qs;
+
+        const float d0 = GGML_FP16_TO_FP32(x[2*i + 0].d);
+        const float m0 = GGML_FP16_TO_FP32(x[2*i + 0].m);
+        const float d1 = GGML_FP16_TO_FP32(x[2*i + 1].d);
+        const float m1 = GGML_FP16_TO_FP32(x[2*i + 1].m);
+
+        int sy_0 = 0;
+        int sy_1 = 0;
+
+        int sxy_0 = 0;
+        int sxy_1 = 0;
+
+        for (int j = 0; j < QK8_0/4; j++) {
+            const uint8_t v0 = x0[j];
+            const uint8_t v1 = x1[j];
+
+            const int x0_0 = v0 & 0xf;
+            const int x1_0 = v0 >> 4;
+
+            const int x0_1 = v1 & 0xf;
+            const int x1_1 = v1 >> 4;
+
+            const int y0_0 = y0[2*j + 0];
+            const int y1_0 = y0[2*j + 1];
+
+            const int y0_1 = y0[2*(j + QK8_0/4) + 0];
+            const int y1_1 = y0[2*(j + QK8_0/4) + 1];
+
+            sy_0 += y0_0 + y1_0;
+            sy_1 += y0_1 + y1_1;
+
+            sxy_0 += x0_0*y0_0 + x1_0*y1_0;
+            sxy_1 += x0_1*y0_1 + x1_1*y1_1;
+        }
+
+        sumf += (d0*sxy_0 + m0*sy_0)*y[i].d;
+        sumf += (d1*sxy_1 + m1*sy_1)*y[i].d;
+    }
+#endif
+
+    *s = sumf;
+}
+
+
 // compute GGML_VEC_DOT_UNROLL dot products at once
 // xs - x row stride in bytes
 inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * restrict s, void * restrict xv, ggml_fp16_t * restrict y) {
@@ -3056,12 +3321,13 @@ static const int GGML_BLCK_SIZE[GGML_TYPE_COUNT] = {
     [GGML_TYPE_Q4_0] = QK4_0,
     [GGML_TYPE_Q4_1] = QK4_1,
     [GGML_TYPE_Q4_2] = QK4_2,
+    [GGML_TYPE_Q4_3] = QK4_3,
     [GGML_TYPE_Q8_0] = QK8_0,
     [GGML_TYPE_I8]   = 1,
     [GGML_TYPE_I16]  = 1,
     [GGML_TYPE_I32]  = 1,
 };
-static_assert(GGML_TYPE_COUNT == 9, "GGML_BLCK_SIZE is outdated");
+static_assert(GGML_TYPE_COUNT == 10, "GGML_BLCK_SIZE is outdated");
 
 static const size_t GGML_TYPE_SIZE[GGML_TYPE_COUNT] = {
     [GGML_TYPE_F32]  = sizeof(float),
@@ -3069,12 +3335,13 @@ static const size_t GGML_TYPE_SIZE[GGML_TYPE_COUNT] = {
     [GGML_TYPE_Q4_0] = sizeof(block_q4_0),
     [GGML_TYPE_Q4_1] = sizeof(block_q4_1),
     [GGML_TYPE_Q4_2] = sizeof(block_q4_2),
+    [GGML_TYPE_Q4_3] = sizeof(block_q4_3),
     [GGML_TYPE_Q8_0] = sizeof(block_q8_0),
     [GGML_TYPE_I8]   = sizeof(int8_t),
     [GGML_TYPE_I16]  = sizeof(int16_t),
     [GGML_TYPE_I32]  = sizeof(int32_t),
 };
-static_assert(GGML_TYPE_COUNT == 9, "GGML_TYPE_SIZE is outdated");
+static_assert(GGML_TYPE_COUNT == 10, "GGML_TYPE_SIZE is outdated");
 
 
 static const char * GGML_TYPE_NAME[GGML_TYPE_COUNT] = {
@@ -3083,12 +3350,13 @@ static const char * GGML_TYPE_NAME[GGML_TYPE_COUNT] = {
     [GGML_TYPE_Q4_0] = "q4_0",
     [GGML_TYPE_Q4_1] = "q4_1",
     [GGML_TYPE_Q4_2] = "q4_2",
+    [GGML_TYPE_Q4_3] = "q4_3",
     [GGML_TYPE_Q8_0] = "q8_0",
     [GGML_TYPE_I8]   = "i8",
     [GGML_TYPE_I16]  = "i16",
     [GGML_TYPE_I32]  = "i32",
 };
-static_assert(GGML_TYPE_COUNT == 9, "GGML_TYPE_NAME is outdated");
+static_assert(GGML_TYPE_COUNT == 10, "GGML_TYPE_NAME is outdated");
 
 static bool GGML_IS_QUANTIZED[GGML_TYPE_COUNT] = {
     [GGML_TYPE_F32]  = false,
@@ -3096,12 +3364,13 @@ static bool GGML_IS_QUANTIZED[GGML_TYPE_COUNT] = {
     [GGML_TYPE_Q4_0] = true,
     [GGML_TYPE_Q4_1] = true,
     [GGML_TYPE_Q4_2] = true,
+    [GGML_TYPE_Q4_3] = true,
     [GGML_TYPE_Q8_0] = true,
     [GGML_TYPE_I8]   = false,
     [GGML_TYPE_I16]  = false,
     [GGML_TYPE_I32]  = false,
 };
-static_assert(GGML_TYPE_COUNT == 9, "GGML_IS_QUANTIZED is outdated");
+static_assert(GGML_TYPE_COUNT == 10, "GGML_IS_QUANTIZED is outdated");
 
 static const char * GGML_OP_LABEL[GGML_OP_COUNT] = {
     "NONE",
@@ -3363,7 +3632,7 @@ static inline bool ggml_can_mul_mat(const struct ggml_tensor * t0, const struct
         (t0->ne[3] == t1->ne[3]);
 }
 
-static inline bool ggml_is_quantized(enum ggml_type type) {
+bool ggml_is_quantized(enum ggml_type type) {
     return GGML_IS_QUANTIZED[type];
 }
 
@@ -5899,7 +6168,6 @@ static void ggml_compute_forward_dup_f32(
                 i10 += ne00 * ir0;
                 while (i10 >= ne0) {
                     i10 -= ne0;
-                    i11++;
                     if (++i11 == ne1) {
                         i11 = 0;
                         if (++i12 == ne2) {
@@ -6313,6 +6581,7 @@ static void ggml_compute_forward_add(
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q4_2:
+        case GGML_TYPE_Q4_3:
             {
                 ggml_compute_forward_add_q_f32(params, src0, src1, dst);
             } break;
@@ -7952,6 +8221,7 @@ static void ggml_compute_forward_mul_mat(
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q4_2:
+        case GGML_TYPE_Q4_3:
         case GGML_TYPE_Q8_0:
             {
                 ggml_compute_forward_mul_mat_q_f32(params, src0, src1, dst);
@@ -7969,34 +8239,6 @@ static void ggml_compute_forward_mul_mat(
                 GGML_ASSERT(false);
             } break;
     }
-
-#if 0
-    if (src0->type == GGML_TYPE_F16 || src0->type == GGML_TYPE_Q4_1) {
-        static int first = 8;
-        printf("src0: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", src0->ne[0], src0->ne[1], src0->ne[2]);
-        printf("src1: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", src1->ne[0], src1->ne[1], src1->ne[2]);
-        printf("dst:  ne0 = %5d, ne1 = %5d, ne2 = %5d\n", dst->ne[0], dst->ne[1], dst->ne[2]);
-        if (first) {
-            --first;
-        } else {
-            for (int k = 0; k < dst->ne[1]; ++k) {
-                for (int j = 0; j < dst->ne[0]/16; ++j) {
-                    for (int i = 0; i < 16; ++i) {
-                        printf("%8.4f ", ((float *) dst->data)[k*dst->ne[0] + j*16 + i]);
-                    }
-                    printf("\n");
-                }
-                printf("\n");
-            }
-            printf("\n");
-            exit(0);
-        }
-    } else {
-        printf("aaaa src0: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", src0->ne[0], src0->ne[1], src0->ne[2]);
-        printf("aaaa src1: ne0 = %5d, ne1 = %5d, ne2 = %5d\n", src1->ne[0], src1->ne[1], src1->ne[2]);
-        printf("aaaa dst:  ne0 = %5d, ne1 = %5d, ne2 = %5d\n", dst->ne[0], dst->ne[1], dst->ne[2]);
-    }
-#endif
 }
 
 // ggml_compute_forward_scale
@@ -8208,6 +8450,7 @@ static void ggml_compute_forward_get_rows(
         case GGML_TYPE_Q4_0:
         case GGML_TYPE_Q4_1:
         case GGML_TYPE_Q4_2:
+        case GGML_TYPE_Q4_3:
         case GGML_TYPE_Q8_0:
             {
                 ggml_compute_forward_get_rows_q(params, src0, src1, dst);
@@ -8438,9 +8681,11 @@ static void ggml_compute_forward_rope_f32(
 
     const float theta_scale = powf(10000.0, -2.0f/n_dims);
 
+    const bool is_neox = mode & 2;
+
     for (int64_t i3 = 0; i3 < ne3; i3++) {
-        for (int64_t i2 = (mode == 0 ? 0 : n_past); i2 < ne2; i2++) {
-            const int p = (mode == 0 ? n_past + i2 : i2);
+        for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) {
+            const int p = ((mode & 1) == 0 ? n_past + i2 : i2);
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
@@ -8453,14 +8698,25 @@ static void ggml_compute_forward_rope_f32(
 
                     theta *= theta_scale;
 
-                    const float * const src = (float *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-                          float * dst_data  = (float *)((char *)  dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+                    if (!is_neox) {
+                        const float * const src = (float *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+                              float * dst_data  = (float *)((char *)  dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
 
-                    const float x0 = src[0];
-                    const float x1 = src[1];
+                        const float x0 = src[0];
+                        const float x1 = src[1];
 
-                    dst_data[0] = x0*cos_theta - x1*sin_theta;
-                    dst_data[1] = x0*sin_theta + x1*cos_theta;
+                        dst_data[0] = x0*cos_theta - x1*sin_theta;
+                        dst_data[1] = x0*sin_theta + x1*cos_theta;
+                    } else {
+                        const float * const src = (float *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0);
+                              float * dst_data  = (float *)((char *)  dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0);
+
+                        const float x0 = src[0];
+                        const float x1 = src[n_dims/2];
+
+                        dst_data[0]        = x0*cos_theta - x1*sin_theta;
+                        dst_data[n_dims/2] = x0*sin_theta + x1*cos_theta;
+                    }
                 }
             }
         }
@@ -8515,9 +8771,11 @@ static void ggml_compute_forward_rope_f16(
 
     const float theta_scale = powf(10000.0, -2.0f/n_dims);
 
+    const bool is_neox = mode & 2;
+
     for (int64_t i3 = 0; i3 < ne3; i3++) {
-        for (int64_t i2 = (mode == 0 ? 0 : n_past); i2 < ne2; i2++) {
-            const int p = (mode == 0 ? n_past + i2 : i2);
+        for (int64_t i2 = ((mode & 1) == 0 ? 0 : n_past); i2 < ne2; i2++) {
+            const int p = ((mode & 1) == 0 ? n_past + i2 : i2);
             for (int64_t i1 = 0; i1 < ne1; i1++) {
                 if (ir++ < ir0) continue;
                 if (ir   > ir1) break;
@@ -8530,14 +8788,25 @@ static void ggml_compute_forward_rope_f16(
 
                     theta *= theta_scale;
 
-                    const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
-                          ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+                    if (!is_neox) {
+                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
+                              ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3 + i2*nb2 + i1*nb1 + i0*nb0);
 
-                    const float x0 = GGML_FP16_TO_FP32(src[0]);
-                    const float x1 = GGML_FP16_TO_FP32(src[1]);
+                        const float x0 = GGML_FP16_TO_FP32(src[0]);
+                        const float x1 = GGML_FP16_TO_FP32(src[1]);
 
-                    dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
-                    dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                        dst_data[0] = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                        dst_data[1] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                    } else {
+                        const ggml_fp16_t * const src = (ggml_fp16_t *)((char *) src0->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0);
+                              ggml_fp16_t * dst_data  = (ggml_fp16_t *)((char *)  dst->data + i3*nb3 + i2*nb2 + i1*nb1 + (i0/2)*nb0);
+
+                        const float x0 = GGML_FP16_TO_FP32(src[0]);
+                        const float x1 = GGML_FP16_TO_FP32(src[n_dims/2]);
+
+                        dst_data[0]        = GGML_FP32_TO_FP16(x0*cos_theta - x1*sin_theta);
+                        dst_data[n_dims/2] = GGML_FP32_TO_FP16(x0*sin_theta + x1*cos_theta);
+                    }
                 }
             }
         }
@@ -11947,6 +12216,62 @@ size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t *
     return (n/QK4_2*sizeof(block_q4_2));
 }
 
+size_t ggml_quantize_q4_3(const float * src, void * dst, int n, int k, int64_t * hist) {
+    assert(k % QK4_3 == 0);
+    const int nb = k / QK4_3;
+
+    for (int j = 0; j < n; j += k) {
+        block_q4_3 * restrict y = (block_q4_3 *)dst + j/QK4_3;
+
+        quantize_row_q4_3_reference(src + j, y, k);
+
+        for (int i = 0; i < nb; i++) {
+            for (int l = 0; l < QK4_3; l += 2) {
+                const uint8_t vi0 = y[i].qs[l/2] & 0xF;
+                const uint8_t vi1 = y[i].qs[l/2] >> 4;
+
+                hist[vi0]++;
+                hist[vi1]++;
+            }
+        }
+    }
+
+    return (n/QK4_3*sizeof(block_q4_3));
+}
+
+size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist) {
+    size_t result = 0;
+    switch (type) {
+        case GGML_TYPE_Q4_0:
+            {
+                GGML_ASSERT(start % QK4_0 == 0);
+                block_q4_0 * block = (block_q4_0*)dst + start / QK4_0;
+                result = ggml_quantize_q4_0(src + start, block, n, n, hist);
+            } break;
+        case GGML_TYPE_Q4_1:
+            {
+                GGML_ASSERT(start % QK4_1 == 0);
+                block_q4_1 * block = (block_q4_1*)dst + start / QK4_1;
+                result = ggml_quantize_q4_1(src + start, block, n, n, hist);
+            } break;
+        case GGML_TYPE_Q4_2:
+            {
+                GGML_ASSERT(start % QK4_2 == 0);
+                block_q4_2 * block = (block_q4_2*)dst + start / QK4_2;
+                result = ggml_quantize_q4_2(src + start, block, n, n, hist);
+            } break;
+        case GGML_TYPE_Q4_3:
+            {
+                GGML_ASSERT(start % QK4_3 == 0);
+                block_q4_3 * block = (block_q4_3*)dst + start / QK4_3;
+                result = ggml_quantize_q4_3(src + start, block, n, n, hist);
+            } break;
+        default:
+            assert(false);
+    }
+    return result;
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 
 int ggml_cpu_has_avx(void) {

ggml.h

@@ -205,7 +205,8 @@ enum ggml_type {
     GGML_TYPE_Q4_0 = 2,
     GGML_TYPE_Q4_1 = 3,
     GGML_TYPE_Q4_2 = 4,
-    GGML_TYPE_Q8_0 = 5,
+    GGML_TYPE_Q4_3 = 5,
+    GGML_TYPE_Q8_0 = 6,
     GGML_TYPE_I8,
     GGML_TYPE_I16,
     GGML_TYPE_I32,
@@ -360,6 +361,8 @@ const char * ggml_type_name(enum ggml_type type);
 
 size_t ggml_element_size(const struct ggml_tensor * tensor);
 
+bool ggml_is_quantized(enum ggml_type type);
+
 struct ggml_context * ggml_init(struct ggml_init_params params);
 void ggml_free(struct ggml_context * ctx);
 
@@ -627,7 +630,8 @@ struct ggml_tensor * ggml_soft_max(
 
 // rotary position embedding
 // in-place, returns view(a)
-// if mode == 1, skip n_past elements
+// if mode & 1 == 1, skip n_past elements
+// if mode & 2 == 1, GPT-NeoX style
 // TODO: avoid creating a new tensor every time
 struct ggml_tensor * ggml_rope(
         struct ggml_context * ctx,
@@ -808,6 +812,9 @@ enum ggml_opt_result ggml_opt(
 size_t ggml_quantize_q4_0(const float * src, void * dst, int n, int k, int64_t * hist);
 size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t * hist);
 size_t ggml_quantize_q4_2(const float * src, void * dst, int n, int k, int64_t * hist);
+size_t ggml_quantize_q4_3(const float * src, void * dst, int n, int k, int64_t * hist);
+
+size_t ggml_quantize_chunk(enum ggml_type type, const float * src, void * dst, int start, int n, int64_t * hist);
 
 //
 // system info

llama.cpp

@@ -24,6 +24,9 @@
 #include <memory>
 #include <algorithm>
 #include <initializer_list>
+#include <thread>
+#include <atomic>
+#include <mutex>
 
 #define LLAMA_USE_SCRATCH
 #define LLAMA_MAX_SCRATCH_BUFFERS 16
@@ -479,6 +482,7 @@ struct llama_file_loader {
                 case GGML_TYPE_Q4_0:
                 case GGML_TYPE_Q4_1:
                 case GGML_TYPE_Q4_2:
+                case GGML_TYPE_Q4_3:
                     break;
                 default: {
                     throw format("unrecognized tensor type %u\n", shard.type);
@@ -552,6 +556,7 @@ struct llama_file_saver {
             case GGML_TYPE_Q4_0:
             case GGML_TYPE_Q4_1:
             case GGML_TYPE_Q4_2:
+            case GGML_TYPE_Q4_3:
                 break;
             default: LLAMA_ASSERT(false);
         }
@@ -841,6 +846,7 @@ static const char *llama_ftype_name(enum llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16:
                                       return "mostly Q4_1, some F16";
         case LLAMA_FTYPE_MOSTLY_Q4_2: return "mostly Q4_2";
+        case LLAMA_FTYPE_MOSTLY_Q4_3: return "mostly Q4_3";
         default:                      return "unknown, may not work";
     }
 }
@@ -1569,15 +1575,20 @@ static llama_vocab::id llama_sample_top_p_top_k(
 // quantization
 //
 
-static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, enum llama_ftype ftype) {
+static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, enum llama_ftype ftype, int nthread) {
     ggml_type quantized_type;
     switch (ftype) {
         case LLAMA_FTYPE_MOSTLY_Q4_0: quantized_type = GGML_TYPE_Q4_0; break;
         case LLAMA_FTYPE_MOSTLY_Q4_1: quantized_type = GGML_TYPE_Q4_1; break;
         case LLAMA_FTYPE_MOSTLY_Q4_2: quantized_type = GGML_TYPE_Q4_2; break;
+        case LLAMA_FTYPE_MOSTLY_Q4_3: quantized_type = GGML_TYPE_Q4_3; break;
         default: throw format("invalid output file type %d\n", ftype);
     };
 
+    if (nthread <= 0) {
+        nthread = std::thread::hardware_concurrency();
+    }
+
     std::unique_ptr<llama_model_loader> model_loader(new llama_model_loader(fname_inp.c_str(), /*use_mmap*/ false,
                                                                             /*vocab_only*/ false));
     llama_file_saver file_saver(fname_out.c_str(), model_loader->file_loaders.at(0).get(), ftype);
@@ -1586,6 +1597,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
     size_t total_size_new = 0;
     std::vector<int64_t> hist_all(1 << 4, 0);
 
+    std::vector<std::thread> workers;
+    std::mutex mutex;
+
     size_t idx = 0;
     for (llama_load_tensor & tensor : model_loader->tensors_map.tensors) {
         llama_buffer read_data;
@@ -1604,6 +1618,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         // quantize only 2D tensors
         quantize &= (tensor.ne.size() == 2);
 
+        // uncomment this to keep the output layer in FP16
+        //if (tensor.name == "output.weight") {
+        //    quantize = false;
+        //}
+
         enum ggml_type new_type;
         void * new_data;
         size_t new_size;
@@ -1639,21 +1658,37 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
             new_data = work.addr;
             std::vector<int64_t> hist_cur(1 << 4, 0);
 
-            switch (new_type) {
-                case GGML_TYPE_Q4_0:
-                    {
-                        new_size = ggml_quantize_q4_0(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data());
-                    } break;
-                case GGML_TYPE_Q4_1:
-                    {
-                        new_size = ggml_quantize_q4_1(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data());
-                    } break;
-                case GGML_TYPE_Q4_2:
-                    {
-                        new_size = ggml_quantize_q4_2(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data());
-                    } break;
-                default:
-                    LLAMA_ASSERT(false);
+            int chunk_size = 32 * 512;
+            const int nchunk = (nelements + chunk_size - 1)/chunk_size;
+            const int nthread_use = nthread > 1 ? std::max(1, std::min(nthread, nchunk)) : 1;
+            if (nthread_use < 2) {
+                new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, nelements, hist_cur.data());
+            } else {
+                size_t counter = 0;
+                new_size = 0;
+                auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, nelements, chunk_size] () {
+                    std::vector<int64_t> local_hist;
+                    size_t local_size = 0;
+                    while (true) {
+                        std::unique_lock<std::mutex> lock(mutex);
+                        size_t first = counter; counter += chunk_size;
+                        if (first >= nelements) {
+                            if (!local_hist.empty()) {
+                                for (int j=0; j<int(local_hist.size()); ++j) hist_cur[j] += local_hist[j];
+                                new_size += local_size;
+                            }
+                            break;
+                        }
+                        lock.unlock();
+                        size_t last = std::min(nelements, first + chunk_size);
+                        if (local_hist.empty()) local_hist.resize(hist_cur.size(), 0);
+                        local_size += ggml_quantize_chunk(new_type, f32_data, new_data, first, last - first, local_hist.data());
+                    }
+                };
+                if (int(workers.size()) < nthread_use - 1) workers.resize(nthread_use - 1);
+                for (int it = 0; it < nthread_use - 1; ++it) workers[it] = std::thread(compute);
+                compute();
+                for (int it = 0; it < nthread_use - 1; ++it) workers[it].join();
             }
 
             printf("size = %8.2f MB -> %8.2f MB | hist: ", tensor.size/1024.0/1024.0, new_size/1024.0/1024.0);
@@ -1775,9 +1810,10 @@ void llama_free(struct llama_context * ctx) {
 int llama_model_quantize(
         const char * fname_inp,
         const char * fname_out,
-  enum llama_ftype   ftype) {
+  enum llama_ftype   ftype,
+        int          nthread) {
     try {
-        llama_model_quantize_internal(fname_inp, fname_out, ftype);
+        llama_model_quantize_internal(fname_inp, fname_out, ftype, nthread);
         return 0;
     } catch (const std::string & err) {
         fprintf(stderr, "%s: failed to quantize: %s\n", __func__, err.c_str());
@@ -1963,7 +1999,7 @@ int llama_apply_lora_from_file_internal(struct llama_context * ctx, const char *
                 base_t = dest_t;
             }
 
-            if (base_t->type == GGML_TYPE_Q4_0 || base_t->type == GGML_TYPE_Q4_1 || base_t->type == GGML_TYPE_Q4_2) {
+            if (ggml_is_quantized(base_t->type)) {
                 if (!warned) {
                     fprintf(stderr, "%s: warning: using a lora adapter with a quantized model may result in poor quality, "
                                     "use a f16 or f32 base model with --lora-base\n", __func__);
@@ -2056,7 +2092,11 @@ void llama_set_kv_cache(
                          int   n_token_count) {
     // Make sure we have the same kv cache setup
     LLAMA_ASSERT(ctx->model.kv_self.buf.size == n_size);
+    void * k_data = ctx->model.kv_self.k->data; // remember data pointers
+    void * v_data = ctx->model.kv_self.v->data; // because their value is stored in buf and overwritten by memcpy
     memcpy(ctx->model.kv_self.buf.addr, kv_cache, n_size);
+    ctx->model.kv_self.k->data = k_data; // restore correct data pointers
+    ctx->model.kv_self.v->data = v_data;
     ctx->model.kv_self.n = n_token_count;
 }
 

llama.h

@@ -73,6 +73,7 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q4_1 = 3,  // except 1d tensors
         LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
         LLAMA_FTYPE_MOSTLY_Q4_2 = 5,  // except 1d tensors
+        LLAMA_FTYPE_MOSTLY_Q4_3 = 6,  // except 1d tensors
     };
 
     LLAMA_API struct llama_context_params llama_context_default_params();
@@ -92,10 +93,12 @@ extern "C" {
 
     // TODO: not great API - very likely to change
     // Returns 0 on success
+    // nthread - how many threads to use. If <=0, will use std::thread::hardware_concurrency(), else the number given
     LLAMA_API int llama_model_quantize(
             const char * fname_inp,
             const char * fname_out,
-      enum llama_ftype   ftype);
+      enum llama_ftype   ftype,
+            int          nthread);
 
     // Apply a LoRA adapter to a loaded model
     // path_base_model is the path to a higher quality model to use as a base for

pocs/vdot/CMakeLists.txt

@@ -2,3 +2,8 @@ set(TARGET vdot)
 add_executable(${TARGET} vdot.cpp)
 target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
 target_compile_features(${TARGET} PRIVATE cxx_std_11)
+
+set(TARGET q8dot)
+add_executable(${TARGET} q8dot.cpp)
+target_link_libraries(${TARGET} PRIVATE common llama ${CMAKE_THREAD_LIBS_INIT})
+target_compile_features(${TARGET} PRIVATE cxx_std_11)

pocs/vdot/q8dot.cpp

@@ -0,0 +1,172 @@
+#include <cstdio>
+#include <type_traits>
+#include <vector>
+#include <random>
+#include <chrono>
+#include <cstdlib>
+#include <cmath>
+#include <cassert>
+#include <cstring>
+#include <array>
+#include <type_traits>
+
+#include <ggml.h>
+
+constexpr int kVecSize = 1 << 16;
+
+// Copy-pasted from ggml.c
+#define QK4_0 32
+typedef struct {
+    float   d;          // delta
+    uint8_t qs[QK4_0 / 2];  // nibbles / quants
+} block_q4_0;
+static_assert(sizeof(block_q4_0) == sizeof(float) + QK4_0 / 2, "wrong q4_0 block size/padding");
+
+#define QK4_1 32
+typedef struct {
+    float   d;          // delta
+    float   m;          // min
+    uint8_t qs[QK4_1 / 2];  // nibbles / quants
+} block_q4_1;
+static_assert(sizeof(block_q4_1) == sizeof(float) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
+
+// Copy-pasted from ggml.c
+#define QK8_0 32
+typedef struct {
+    float   d;          // delta
+    float   s;          // d * sum(qs[i])
+    int8_t  qs[QK8_0];  // quants
+} block_q8_0;
+static_assert(sizeof(block_q8_0) == 2*sizeof(float) + QK8_0, "wrong q8_0 block size/padding");
+
+static_assert(QK4_1 == QK8_0, "QK4_1 and QK8_0 must be the same");
+static_assert(QK4_0 == QK8_0, "QK4_0 and QK8_0 must be the same");
+
+template <typename T>
+void fillQ4blocks(std::vector<T>& blocks, std::mt19937& rndm) {
+    for (auto& b : blocks) {
+        b.d = 1;
+        for (int i=0; i<QK4_1/2; ++i) {
+            uint8_t v1 = rndm() >> 28;
+            uint8_t v2 = rndm() >> 28;
+            b.qs[i] = v1 | (v2 << 4);
+        }
+    }
+}
+
+void fillQ80blocks(std::vector<block_q8_0>& blocks, std::mt19937& rndm) {
+    for (auto& b : blocks) {
+        b.d = 1;
+        int sum = 0;
+        for (int i=0; i<QK8_0; ++i) {
+            b.qs[i] = (rndm() >> 24) - 128;
+            sum += b.qs[i];
+        }
+        b.s = b.d * sum;
+    }
+}
+
+float simpleDot(const block_q4_0& x, const block_q8_0& y) {
+    int s1 = 0; //, s2 = 0;
+    for (int i=0; i<QK4_1/2; i+=2) {
+        int v1 = x.qs[i+0] & 0xf;
+        int v2 = x.qs[i+0] >> 4;
+        int v3 = x.qs[i+1] & 0xf;
+        int v4 = x.qs[i+1] >> 4;
+        int j = 2*i;
+        s1 += v1*y.qs[j] + v2*y.qs[j+1] + v3*y.qs[j+2] + v4*y.qs[j+3];
+        //s2 += y.qs[j] + y.qs[j+1] + y.qs[j+2] + y.qs[j+3];
+    }
+    return y.d * x.d * s1 - 8 * x.d * y.s;
+    //return y.d * x.d * (s1 - 8 * s2);
+}
+
+float simpleDot(const block_q4_1& x, const block_q8_0& y) {
+    int s1 = 0; //, s2 = 0;
+    for (int i=0; i<QK4_1/2; i+=2) {
+        int v1 = x.qs[i+0] & 0xf;
+        int v2 = x.qs[i+0] >> 4;
+        int v3 = x.qs[i+1] & 0xf;
+        int v4 = x.qs[i+1] >> 4;
+        int j = 2*i;
+        s1 += v1*y.qs[j] + v2*y.qs[j+1] + v3*y.qs[j+2] + v4*y.qs[j+3];
+        //s2 += y.qs[j] + y.qs[j+1] + y.qs[j+2] + y.qs[j+3];
+    }
+    return y.d * x.d * s1 + y.s * x.m;
+    //return y.d * (x.d * s1 + x.m * s2);
+}
+
+struct Stat {
+    double sum = 0, sumt = 0, sumt2 = 0, maxt = 0;
+    int nloop = 0;
+    void addResult(double s, double t) {
+        sum += s;
+        sumt += t; sumt2 += t*t; maxt = std::max(maxt, t);
+        ++nloop;
+    }
+    void reportResult(const char* title) const {
+        if (nloop < 1) {
+            printf("%s(%s): no result\n",__func__,title);
+            return;
+        }
+        printf("============ %s\n",title);
+        printf("<dot> = %g\n",sum/nloop);
+        auto t = sumt/nloop, dt = sumt2/nloop - t*t;
+        if (dt > 0) dt = sqrt(dt);
+        printf("<time> = %g +/- %g us. Max. time = %g us.\n",t,dt,maxt);
+    }
+};
+
+
+int main(int argc, char** argv) {
+
+    int nloop = argc > 1 ? atoi(argv[1]) : 10;
+    int type  = argc > 2 ? atoi(argv[2]) : 1;
+
+    std::mt19937 rndm(1234);
+
+    std::vector<block_q4_1> x41;
+    std::vector<block_q4_0> x40;
+    std::vector<block_q8_0> y(kVecSize);
+    if (type == 0) x40.resize(kVecSize);
+    else {
+        x41.resize(kVecSize);
+        for (auto& b : x41) b.m = 1;
+    }
+
+    auto ggml_type = type == 0 ? GGML_TYPE_Q4_0 : GGML_TYPE_Q4_1;
+
+    auto funcs = ggml_internal_get_quantize_fn(ggml_type);
+
+    Stat simple, ggml;
+
+    for (int iloop=0; iloop<nloop; ++iloop) {
+
+        if (type == 0) fillQ4blocks(x40, rndm);
+        else fillQ4blocks(x41, rndm);
+        fillQ80blocks(y, rndm);
+
+        auto t1 = std::chrono::high_resolution_clock::now();
+        double s = 0;
+        if (type == 0) for (int i=0; i<kVecSize; ++i) s += simpleDot(x40[i], y[i]);
+        else for (int i=0; i<kVecSize; ++i) s += simpleDot(x41[i], y[i]);
+        auto t2 = std::chrono::high_resolution_clock::now();
+        auto t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
+        if (iloop > 3) simple.addResult(s, t);
+
+        t1 = std::chrono::high_resolution_clock::now();
+        float fs;
+        if (type == 0) funcs.vec_dot_q(kVecSize * QK4_1, &fs, x40.data(), y.data());
+        else funcs.vec_dot_q(kVecSize * QK4_1, &fs, x41.data(), y.data());
+        t2 = std::chrono::high_resolution_clock::now();
+        t = 1e-3*std::chrono::duration_cast<std::chrono::nanoseconds>(t2-t1).count();
+        if (iloop > 3) ggml.addResult(fs, t);
+
+    }
+
+    // Report the time (and the average of the dot products so the compiler does not come up with the idea
+    // of optimizing away the function calls after figuring that the result is not used).
+    simple.reportResult("Simple");
+    ggml.reportResult("ggml");
+    return 0;
+}