llama_cpp 0.0.5 → 0.0.6

data/ext/llama_cpp/src/ggml.h

@@ -204,7 +204,9 @@ enum ggml_type {
     GGML_TYPE_F16  = 1,
     GGML_TYPE_Q4_0 = 2,
     GGML_TYPE_Q4_1 = 3,
-    GGML_TYPE_Q8_0 = 4,
+    GGML_TYPE_Q4_2 = 4,
+    GGML_TYPE_Q4_3 = 5,
+    GGML_TYPE_Q8_0 = 6,
     GGML_TYPE_I8,
     GGML_TYPE_I16,
     GGML_TYPE_I32,
@@ -359,6 +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);
 
@@ -626,7 +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,
@@ -806,6 +811,10 @@ 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
@@ -823,6 +832,7 @@ int ggml_cpu_has_f16c(void);
 int ggml_cpu_has_fp16_va(void);
 int ggml_cpu_has_wasm_simd(void);
 int ggml_cpu_has_blas(void);
+int ggml_cpu_has_cublas(void);
 int ggml_cpu_has_sse3(void);
 int ggml_cpu_has_vsx(void);
 

data/ext/llama_cpp/src/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
@@ -478,6 +481,8 @@ struct llama_file_loader {
                 case GGML_TYPE_F16:
                 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);
@@ -550,6 +555,8 @@ struct llama_file_saver {
             case GGML_TYPE_F16:
             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);
         }
@@ -838,6 +845,8 @@ static const char *llama_ftype_name(enum llama_ftype ftype) {
         case LLAMA_FTYPE_MOSTLY_Q4_1: return "mostly Q4_1";
         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";
     }
 }
@@ -1066,7 +1075,7 @@ static bool llama_eval_internal(
     // for big prompts, if BLAS is enabled, it is better to use only one thread
     // otherwise, the threads are spin-lock waiting for the BLAS calls and are degrading the performance
     ggml_cgraph gf = {};
-    gf.n_threads = N >= 32 && ggml_cpu_has_blas() ? 1 : n_threads;
+    gf.n_threads = N >= 32 && ggml_cpu_has_blas() && !ggml_cpu_has_cublas() ? 1 : n_threads;
 
     struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
     memcpy(embd->data, tokens, N*ggml_element_size(embd));
@@ -1566,14 +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);
@@ -1582,6 +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;
@@ -1600,6 +1618,11 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
         // quantize only 2D tensors
         quantize &= (tensor.ne.size() == 2);
 
+        // GG: uncomment this to keep the output layer in FP16
+        //if (tensor.name.rfind("output")) {
+        //    quantize = false;
+        //}
+
         enum ggml_type new_type;
         void * new_data;
         size_t new_size;
@@ -1635,17 +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;
-                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);
@@ -1767,9 +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());
@@ -1955,7 +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) {
+            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__);

data/ext/llama_cpp/src/llama.h

@@ -72,6 +72,8 @@ extern "C" {
         LLAMA_FTYPE_MOSTLY_Q4_0 = 2,  // except 1d tensors
         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();
@@ -91,10 +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

data/ext/llama_cpp/src/llama_util.h

@@ -202,7 +202,6 @@ struct llama_mmap {
 
         HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
         DWORD error = GetLastError();
-        CloseHandle(hFile);
 
         if (hMapping == NULL) {
             throw format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str());

data/lib/llama_cpp/version.rb

@@ -3,8 +3,8 @@
 # llama_cpp.rb provides Ruby bindings for the llama.cpp.
 module LLaMACpp
   # The version of llama_cpp.rb you install.
-  VERSION = '0.0.5'
+  VERSION = '0.0.6'
 
   # The version of llama.cpp bundled with llama_cpp.rb.
-  LLAMA_CPP_VERSION = 'master-315a95a'
+  LLAMA_CPP_VERSION = 'master-12b5900'
 end

data/sig/llama_cpp.rbs

@@ -5,6 +5,15 @@ module LLaMACpp
   LLAMA_FILE_MAGIC: String
   LLAMA_FILE_MAGIC_UNVERSIONED: String
 
+  LLAMA_FTYPE_ALL_F32: Integer
+  LLAMA_FTYPE_MOSTLY_F16: Integer
+  LLAMA_FTYPE_MOSTLY_Q4_0: Integer
+  LLAMA_FTYPE_MOSTLY_Q4_1: Integer
+  LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16: Integer
+  LLAMA_FTYPE_MOSTLY_Q4_2: Integer
+  LLAMA_FTYPE_MOSTLY_Q4_3: Integer
+
+  def self?.model_quantize: (input_path: String, output_path: String, ftype: Integer, ?n_threads: Integer) -> void
   def self?.generate: (::LLaMACpp::Context, String, ?n_threads: Integer) -> String
   def self?.print_system_info: () -> void
   def self?.token_bos: () -> Integer
@@ -18,7 +27,7 @@ module LLaMACpp
     def initialize: (model_path: String, params: ::LLaMACpp::ContextParams) -> void
                   | () -> void
     def embeddings: () -> Array[Float]
-    def eval: (tokens: Array[Integer], n_past: Integer, ?n_tokens: Integer, ?n_threads: Integer) -> Qnil
+    def eval: (tokens: Array[Integer], n_past: Integer, ?n_tokens: Integer, ?n_threads: Integer) -> void
     def free: () -> void
     def load: (model_path: String, params: ::LLaMACpp::ContextParams) -> void
     def logits: () -> Array[Float]

metadata

@@ -1,14 +1,14 @@
 --- !ruby/object:Gem::Specification
 name: llama_cpp
 version: !ruby/object:Gem::Version
-  version: 0.0.5
+  version: 0.0.6
 platform: ruby
 authors:
 - yoshoku
 autorequire:
 bindir: exe
 cert_chain: []
-date: 2023-04-20 00:00:00.000000000 Z
+date: 2023-04-22 00:00:00.000000000 Z
 dependencies: []
 description: llama_cpp.rb provides Ruby bindings for the llama.cpp.
 email:
@@ -26,6 +26,7 @@ files:
 - ext/llama_cpp/llama_cpp.cpp
 - ext/llama_cpp/llama_cpp.h
 - ext/llama_cpp/src/LICENSE
+- ext/llama_cpp/src/ggml-cuda.h
 - ext/llama_cpp/src/ggml.c
 - ext/llama_cpp/src/ggml.h
 - ext/llama_cpp/src/llama.cpp