llama : multi-threaded quantization (#1075)

* Multi-threading quantization.

Not much gain for simple quantizations, bit it will be important
for quantizations that require more CPU cycles.

* Multi-threading for quantize-stats

It now does the job in ~14 seconds on my Mac for
Q4_0, Q4_1 and Q4_2. Single-threaded it was taking
more than 2 minutes after adding the more elaborate
version of Q4_2.

* Reviewer comments

* Avoiding compiler confusion

After changing chunk_size to const int as suggested by
@ggerganov, clang and GCC starting to warn me that I don't
need to capture it in the lambda. So, I removed it from the
capture list. But that makes the MSVC build fail. So,
making it a constexpr to make every compiler happy.

* Still fighting with lambda captures in MSVC

---------

Co-authored-by: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Co-authored-by: Georgi Gerganov <ggerganov@gmail.com>

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,
+        std::vector<float> & input_scratch,
-        char *quantized_scratch,
+        std::vector<char> & quantized_scratch,
-        float * output_scratch,
+        std::vector<float> & output_scratch,
-        error_stats & total_error) {
+        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) {
+    float* input_scratch_ptr = nullptr;
-        int64_t chunk_size = std::min(SCRATCH_ELEMENTS, nelements - offset);
+    if (layer->type == GGML_TYPE_F16) {
-
+        if (input_scratch.size() < nelements) input_scratch.resize(nelements);
-        if (layer->type == GGML_TYPE_F16) {
+        input_scratch_ptr = input_scratch.data();
-            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);
-        }
     }
+    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<float> input_scratch;
-    std::vector<char> quantized_scratch(SCRATCH_ELEMENTS*4);
+    std::vector<char> quantized_scratch;
-    std::vector<float> output_scratch(SCRATCH_ELEMENTS);
+    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(),
+                        input_scratch,
-                        quantized_scratch.data(),
+                        quantized_scratch,
-                        output_scratch.data(),
+                        output_scratch,
-                        global_stats
+                        global_stats,
+                        max_thread
                 );
             }
 

examples/quantize/quantize.cpp

@@ -10,8 +10,8 @@
 int main(int argc, char ** argv) {
     ggml_time_init();
 
-    if (argc != 4) {
+    if (argc < 4) {
-        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+        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);
@@ -30,6 +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();
 
@@ -39,7 +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.c

@@ -12189,6 +12189,33 @@ size_t ggml_quantize_q4_3(const float * src, void * dst, int n, int k, int64_t *
     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;
+        default:
+            assert(false);
+    }
+    return result;
+}
+
 ////////////////////////////////////////////////////////////////////////////////
 
 int ggml_cpu_has_avx(void) {

ggml.h

@@ -813,6 +813,8 @@ size_t ggml_quantize_q4_1(const float * src, void * dst, int n, int k, int64_t *
 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
@@ -1572,7 +1575,7 @@ 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;
@@ -1582,6 +1585,10 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         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);
@@ -1590,6 +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;
@@ -1643,25 +1653,37 @@ 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) {
+            int chunk_size = 32 * 512;
-                case GGML_TYPE_Q4_0:
+            const int nchunk = (nelements + chunk_size - 1)/chunk_size;
-                    {
+            const int nthread_use = nthread > 1 ? std::max(1, std::min(nthread, nchunk)) : 1;
-                        new_size = ggml_quantize_q4_0(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data());
+            if (nthread_use < 2) {
-                    } break;
+                new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, nelements, hist_cur.data());
-                case GGML_TYPE_Q4_1:
+            } else {
-                    {
+                size_t counter = 0;
-                        new_size = ggml_quantize_q4_1(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data());
+                new_size = 0;
-                    } break;
+                auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, nelements, chunk_size] () {
-                case GGML_TYPE_Q4_2:
+                    std::vector<int64_t> local_hist;
-                    {
+                    size_t local_size = 0;
-                        new_size = ggml_quantize_q4_2(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data());
+                    while (true) {
-                    } break;
+                        std::unique_lock<std::mutex> lock(mutex);
-                case GGML_TYPE_Q4_3:
+                        size_t first = counter; counter += chunk_size;
-                    {
+                        if (first >= nelements) {
-                        new_size = ggml_quantize_q4_3(f32_data, new_data, nelements, (int) tensor.ne.at(0), hist_cur.data());
+                            if (!local_hist.empty()) {
-                    } break;
+                                for (int j=0; j<int(local_hist.size()); ++j) hist_cur[j] += local_hist[j];
-                default:
+                                new_size += local_size;
-                    LLAMA_ASSERT(false);
+                            }
+                            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);
@@ -1783,9 +1805,10 @@ 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());

llama.h

@@ -93,10 +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