llama : initial working FP16 + 4-bit Q4_0

CMakeLists.txt

@@ -31,7 +31,7 @@ option(GGML_NO_ACCELERATE           "ggml: disable Accelerate framework" OFF)
 # sanitizers
 
 if (GGML_SANITIZE_THREAD)
-    set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS} -fsanitize=thread")
+    set(CMAKE_C_FLAGS   "${CMAKE_C_FLAGS}   -fsanitize=thread")
     set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -fsanitize=thread")
 endif()
 

examples/CMakeLists.txt

@@ -4,3 +4,4 @@ target_include_directories(ggml_utils PUBLIC ${CMAKE_CURRENT_SOURCE_DIR})
 add_subdirectory(gpt-2)
 add_subdirectory(gpt-j)
 add_subdirectory(whisper)
+add_subdirectory(llama)

examples/gpt-2/CMakeLists.txt

@@ -4,3 +4,10 @@
 set(TEST_TARGET gpt-2)
 add_executable(${TEST_TARGET} main.cpp)
 target_link_libraries(${TEST_TARGET} PRIVATE ggml ggml_utils)
+
+#
+# gpt-2-quantize
+
+set(TEST_TARGET gpt-2-quantize)
+add_executable(${TEST_TARGET} quantize.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml ggml_utils)

examples/gpt-2/README.md

@@ -94,7 +94,7 @@ Done! Model '117M' saved in 'models/gpt-2-117M/'
 
 Run the convert-ckpt-to-ggml.py script to convert the model to ggml format.
 
-  python /Users/john/ggml/examples/gpt-2/convert-ckpt-to-ggml.py models/gpt-2-117M/
+  python /Users/john/ggml/examples/gpt-2/convert-ckpt-to-ggml.py models/gpt-2-117M/ 1
 
 ```
 

examples/gpt-2/convert-ckpt-to-ggml.py

@@ -45,8 +45,18 @@ def bytes_to_unicode():
     cs = [chr(n) for n in cs]
     return dict(zip(bs, cs))
 
-if len(sys.argv) < 2:
-    print("Usage: convert-ckpt-to-ggml.py dir-model [use-f32]\n")
+# helper method to convert a numpy array to different float types
+def convert_to_ftype(data, ftype):
+    # fp16
+    if ftype == 1:
+        return data.astype(np.float16)
+
+    assert False, "Invalid ftype: " + str(ftype)
+
+if len(sys.argv) < 3:
+    print("Usage: convert-ckpt-to-ggml.py dir-model ftype\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
     sys.exit(1)
 
 # output in the same directory as the model
@@ -59,11 +69,20 @@ with open(dir_model + "/encoder.json", "r") as f:
 with open(dir_model + "/hparams.json", "r") as f:
     hparams = json.load(f)
 
-# use 16-bit or 32-bit floats
-use_f16 = True
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
 if len(sys.argv) > 2:
-    use_f16 = False
-    fname_out = sys.argv[1] + "/ggml-model-f32.bin"
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+    fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
 
 list_vars = tf.train.list_variables(dir_model)
 
@@ -75,7 +94,7 @@ fout.write(struct.pack("i", hparams["n_ctx"]))
 fout.write(struct.pack("i", hparams["n_embd"]))
 fout.write(struct.pack("i", hparams["n_head"]))
 fout.write(struct.pack("i", hparams["n_layer"]))
-fout.write(struct.pack("i", use_f16))
+fout.write(struct.pack("i", ftype))
 
 byte_encoder = bytes_to_unicode()
 byte_decoder = {v:k for k, v in byte_encoder.items()}
@@ -93,9 +112,22 @@ for name, shape in list_vars:
     data = tf.train.load_variable(dir_model, name).squeeze()
     n_dims = len(data.shape);
 
-    # ftype == 0 -> float32, ftype == 1 -> float16
-    ftype = 0;
-    if use_f16:
+    # for efficiency - transpose the projection matrices
+    # "model/h.*/attn/c_attn/w"
+    # "model/h.*/attn/c_proj/w"
+    # "model/h.*/mlp/c_fc/w"
+    # "model/h.*/mlp/c_proj/w"
+    if name[-14:] == "/attn/c_attn/w" or \
+       name[-14:] == "/attn/c_proj/w" or \
+       name[-11:] == "/mlp/c_fc/w" or \
+       name[-13:] == "/mlp/c_proj/w":
+        print("  Transposing")
+        data = data.transpose()
+
+    dshape = data.shape
+
+    ftype_cur = 0
+    if ftype != 0:
         # match name:
         #  "model/wte"
         #  "model/h.*/attn/c_attn/w"
@@ -103,24 +135,19 @@ for name, shape in list_vars:
         #  "model/h.*/mlp/c_fc/w"
         #  "model/h.*/mlp/c_proj/w"
         if name == "model/wte" or name[-2:] == "/w":
-            print("  Converting to float16")
-            data = data.astype(np.float16)
-            ftype = 1
+            print("  Converting to " + ftype_str[ftype])
+            data = convert_to_ftype(data, ftype)
+            ftype_cur = ftype
         else:
             print("  Converting to float32")
             data = data.astype(np.float32)
-            ftype = 0
-
-    # for efficiency - transpose the projection matrices
-    if name[-13:] == "/mlp/c_proj/w":
-        print("  Transposing")
-        data = data.transpose()
+            ftype_cur = 0
 
     # header
     str = name.encode('utf-8')
-    fout.write(struct.pack("iii", n_dims, len(str), ftype))
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
     for i in range(n_dims):
-        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+        fout.write(struct.pack("i", dshape[n_dims - 1 - i]))
     fout.write(str);
 
     # data

examples/gpt-2/main.cpp

@@ -128,9 +128,23 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
         }
     }
 
-    // for the big tensors, we have the option to store the data in 16-bit floats
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
     // in order to save memory and also to speed up the computation
-    const ggml_type wtype = model.hparams.f16 ? GGML_TYPE_F16 : GGML_TYPE_F32;
+    ggml_type wtype = GGML_TYPE_COUNT;
+    switch (model.hparams.f16) {
+        case 0: wtype = GGML_TYPE_F32;  break;
+        case 1: wtype = GGML_TYPE_F16;  break;
+        case 2: wtype = GGML_TYPE_Q4_0; break;
+        case 3: wtype = GGML_TYPE_Q4_1; break;
+        default:
+                {
+                    fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n",
+                            __func__, fname.c_str(), model.hparams.f16);
+                    return false;
+                }
+    }
+
+    const ggml_type wtype2 = GGML_TYPE_F32;
 
     auto & ctx = model.ctx;
 
@@ -144,32 +158,32 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
         const int n_ctx   = hparams.n_ctx;
         const int n_vocab = hparams.n_vocab;
 
-        ctx_size += n_embd*ggml_type_size(GGML_TYPE_F32); // ln_f_g
-        ctx_size += n_embd*ggml_type_size(GGML_TYPE_F32); // ln_f_b
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
 
-        ctx_size += n_vocab*n_embd*ggml_type_size(wtype);         // wte
-        ctx_size +=   n_ctx*n_embd*ggml_type_size(GGML_TYPE_F32); // wpe
+        ctx_size += n_vocab*n_embd*ggml_type_sizef(wtype);         // wte
+        ctx_size +=   n_ctx*n_embd*ggml_type_sizef(GGML_TYPE_F32); // wpe
 
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_g
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_b
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
 
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_2_g
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_2_b
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_2_b
 
-        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_size(wtype));         // c_attn_attn_w
-        ctx_size += n_layer*(       3*n_embd*ggml_type_size(GGML_TYPE_F32)); // c_attn_attn_b
+        ctx_size += n_layer*(3*n_embd*n_embd*ggml_type_sizef(wtype));         // c_attn_attn_w
+        ctx_size += n_layer*(       3*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_attn_attn_b
 
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_size(wtype));           // c_attn_proj_w
-        ctx_size += n_layer*(       n_embd*ggml_type_size(GGML_TYPE_F32));   // c_attn_proj_b
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype));           // c_attn_proj_w
+        ctx_size += n_layer*(       n_embd*ggml_type_sizef(GGML_TYPE_F32));   // c_attn_proj_b
 
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_size(wtype));         // c_mlp_fc_w
-        ctx_size += n_layer*(       4*n_embd*ggml_type_size(GGML_TYPE_F32)); // c_mlp_fc_b
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
 
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_size(wtype));         // c_mlp_proj_w
-        ctx_size += n_layer*(         n_embd*ggml_type_size(GGML_TYPE_F32)); // c_mlp_proj_b
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
 
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_size(GGML_TYPE_F32); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_size(GGML_TYPE_F32); // memory_v
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
 
         ctx_size += (6 + 12*n_layer)*256; // object overhead
 
@@ -223,13 +237,13 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
             layer.ln_2_g             = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
             layer.ln_2_b             = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
 
-            layer.c_attn_attn_w      = ggml_new_tensor_2d(ctx, wtype,         3*n_embd, n_embd);
+            layer.c_attn_attn_w      = ggml_new_tensor_2d(ctx, wtype,           n_embd, 3*n_embd);
             layer.c_attn_attn_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*n_embd);
 
             layer.c_attn_proj_w      = ggml_new_tensor_2d(ctx, wtype,           n_embd, n_embd);
             layer.c_attn_proj_b      = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
 
-            layer.c_mlp_fc_w         = ggml_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
+            layer.c_mlp_fc_w         = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
             layer.c_mlp_fc_b         = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
 
             layer.c_mlp_proj_w_trans = ggml_new_tensor_2d(ctx, wtype,         4*n_embd, n_embd);
@@ -319,9 +333,26 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
                 return false;
             }
 
-            const size_t bpe = (ftype == 0) ? sizeof(float) : sizeof(ggml_fp16_t);
+            if (0) {
+                static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ftype_str[ftype], ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
 
-            if (nelements*bpe != ggml_nbytes(tensor)) {
+            size_t bpe = 0;
+
+            switch (ftype) {
+                case 0: bpe = ggml_type_size(GGML_TYPE_F32);  break;
+                case 1: bpe = ggml_type_size(GGML_TYPE_F16);  break;
+                case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break;
+                case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break;
+                default:
+                        {
+                            fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
+                            return false;
+                        }
+            };
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
                 fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
                         __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
                 return false;
@@ -329,7 +360,6 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
 
             fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
 
-            //printf("%24s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
             total_size += ggml_nbytes(tensor);
         }
 
@@ -431,7 +461,7 @@ bool gpt2_eval(
         // [2304, N]
         {
             cur = ggml_mul_mat(ctx0,
-                    ggml_transpose(ctx0, model.layers[il].c_attn_attn_w),
+                    model.layers[il].c_attn_attn_w,
                     cur);
 
             cur = ggml_add(ctx0,
@@ -538,7 +568,7 @@ bool gpt2_eval(
         // [768, N]
         {
             cur = ggml_mul_mat(ctx0,
-                    ggml_transpose(ctx0, model.layers[il].c_attn_proj_w),
+                    model.layers[il].c_attn_proj_w,
                     cur);
 
             cur = ggml_add(ctx0,
@@ -575,7 +605,7 @@ bool gpt2_eval(
             // cur = fc_w*cur + fc_b
             // [3072, N]
             cur = ggml_mul_mat(ctx0,
-                    ggml_transpose(ctx0, model.layers[il].c_mlp_fc_w),
+                    model.layers[il].c_mlp_fc_w,
                     cur);
 
             cur = ggml_add(ctx0,
@@ -705,8 +735,12 @@ int main(int argc, char ** argv) {
 
     params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
 
-    printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
-    printf("\n");
+    printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+    printf("%s: number of tokens in prompt = %zu, first 8 tokens: ", __func__, embd_inp.size());
+    for (int i = 0; i < std::min(8, (int) embd_inp.size()); i++) {
+        printf("%d ", embd_inp[i]);
+    }
+    printf("\n\n");
 
     // submit the input prompt token-by-token
     // this reduces the memory usage during inference, at the cost of a bit of speed at the beginning

examples/gpt-2/quantize.cpp

@@ -0,0 +1,325 @@
+#include "ggml/ggml.h"
+
+#include "utils.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <regex>
+
+// TODO: move somewhere else
+#define QK 32
+
+// default hparams (GPT-2 117M)
+struct gpt2_hparams {
+    int32_t n_vocab = 50257;
+    int32_t n_ctx   = 1024;
+    int32_t n_embd  = 768;
+    int32_t n_head  = 12;
+    int32_t n_layer = 12;
+    int32_t f16     = 1;
+};
+
+// quantize a model
+bool gpt2_model_quantize(const std::string & fname_inp, const std::string & fname_out, int itype) {
+    ggml_type type = GGML_TYPE_Q4_1;
+
+    switch (itype) {
+        case 2: type = GGML_TYPE_Q4_0; break;
+        case 3: type = GGML_TYPE_Q4_1; break;
+        default: fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype); return 1;
+    };
+
+    if (type != GGML_TYPE_Q4_0 && type != GGML_TYPE_Q4_1) {
+        fprintf(stderr, "%s: invalid quantization type %d\n", __func__, type);
+        return false;
+    }
+
+    gpt_vocab vocab;
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *) &magic, sizeof(magic));
+        if (magic != 0x67676d6c) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *) &magic, sizeof(magic));
+    }
+
+    gpt2_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        finp.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        finp.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        finp.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        finp.read((char *) &hparams.f16,     sizeof(hparams.f16));
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: f16     = %d\n", __func__, hparams.f16);
+
+        fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fout.write((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fout.write((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fout.write((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fout.write((char *) &itype,           sizeof(hparams.f16));
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        finp.read ((char *) &n_vocab, sizeof(n_vocab));
+        fout.write((char *) &n_vocab, sizeof(n_vocab));
+
+        if (n_vocab != hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname_inp.c_str(), n_vocab, hparams.n_vocab);
+            return false;
+        }
+
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            finp.read ((char *) &len, sizeof(len));
+            fout.write((char *) &len, sizeof(len));
+
+            word.resize(len);
+            finp.read ((char *) word.data(), len);
+            fout.write((char *) word.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // load weights
+    {
+        size_t total_size_org = 0;
+        size_t total_size_new = 0;
+
+        std::vector<float> work;
+
+        std::vector<uint8_t>     data_u8;
+        std::vector<ggml_fp16_t> data_f16;
+        std::vector<float>       data_f32;
+
+        std::vector<int64_t> hist_all(1 << 4, 0);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ftype;
+
+            finp.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            finp.read(reinterpret_cast<char *>(&length), sizeof(length));
+            finp.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+
+            if (finp.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                finp.read (reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            finp.read (&name[0], length);
+
+            {
+                static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
+                printf("%24s - [%5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ftype_str[ftype]);
+            }
+
+            // regexes of tensor names to be quantized
+            const std::vector<std::string> k_names = {
+                "model/wte",
+                "model/h.*/attn/c_attn/w",
+                "model/h.*/attn/c_proj/w",
+                "model/h.*/mlp/c_fc/w",
+                "model/h.*/mlp/c_proj/w",
+            };
+
+            bool quantize = false;
+            for (const auto & s : k_names) {
+                if (std::regex_match(name, std::regex(s))) {
+                    quantize = true;
+                    break;
+                }
+            }
+
+            if (quantize) {
+                if (ftype != 0 && ftype != 1) {
+                    fprintf(stderr, "%s: unsupported ftype %d for integer quantization\n", __func__, ftype);
+                    return false;
+                }
+
+                if (ftype == 1) {
+                    data_f16.resize(nelements);
+                    finp.read(reinterpret_cast<char *>(data_f16.data()), nelements * sizeof(ggml_fp16_t));
+                    data_f32.resize(nelements);
+                    for (int i = 0; i < nelements; ++i) {
+                        data_f32[i] = ggml_fp16_to_fp32(data_f16[i]);
+                    }
+                } else {
+                    data_f32.resize(nelements);
+                    finp.read(reinterpret_cast<char *>(data_f32.data()), nelements * sizeof(float));
+                }
+
+                ftype = itype;
+            } else {
+                const int bpe = (ftype == 0) ? sizeof(float) : sizeof(uint16_t);
+
+                data_u8.resize(nelements*bpe);
+                finp.read(reinterpret_cast<char *>(data_u8.data()), nelements * bpe);
+            }
+
+            fout.write(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fout.write(reinterpret_cast<char *>(&length), sizeof(length));
+            fout.write(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+            for (int i = 0; i < n_dims; ++i) {
+                fout.write(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+            }
+            fout.write(&name[0], length);
+
+            if (quantize) {
+                printf("quantizing .. ");
+                work.resize(nelements); // for quantization
+
+                size_t cur_size = 0;
+                std::vector<int64_t> hist_cur(1 << 4, 0);
+
+                switch (type) {
+                    case GGML_TYPE_Q4_0:
+                        {
+                            cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], QK, hist_cur.data());
+                        } break;
+                    case GGML_TYPE_Q4_1:
+                        {
+                            cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], QK, hist_cur.data());
+                        } break;
+                    default:
+                        {
+                            fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, type);
+                            return false;
+                        }
+                }
+
+                fout.write(reinterpret_cast<char *>(work.data()), cur_size);
+                total_size_new += cur_size;
+
+                printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
+                for (int i = 0; i < hist_cur.size(); ++i) {
+                    hist_all[i] += hist_cur[i];
+                }
+
+                for (int i = 0; i < hist_cur.size(); ++i) {
+                    printf("%5.3f ", hist_cur[i] / (float)nelements);
+                }
+                printf("\n");
+            } else {
+                printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
+                fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
+                total_size_new += data_u8.size();
+            }
+
+            total_size_org += nelements * sizeof(float);
+        }
+
+        printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
+        printf("%s: quant size  = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);
+
+        {
+            int64_t sum_all = 0;
+            for (int i = 0; i < hist_all.size(); ++i) {
+                sum_all += hist_all[i];
+            }
+
+            printf("%s: hist: ", __func__);
+            for (int i = 0; i < hist_all.size(); ++i) {
+                printf("%5.3f ", hist_all[i] / (float)sum_all);
+            }
+            printf("\n");
+        }
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+// usage:
+//  ./gpt-2-quantize models/gpt-2-117M/ggml-model.bin models/gpt-2-117M/ggml-model-quant.bin type
+//
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+        fprintf(stderr, "  type = 2 - q4_0\n");
+        fprintf(stderr, "  type = 3 - q4_1\n");
+        return 1;
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const int itype = atoi(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!gpt2_model_quantize(fname_inp, fname_out, itype)) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    return 0;
+}

examples/gpt-j/CMakeLists.txt

@@ -4,3 +4,10 @@
 set(TEST_TARGET gpt-j)
 add_executable(${TEST_TARGET} main.cpp)
 target_link_libraries(${TEST_TARGET} PRIVATE ggml ggml_utils)
+
+#
+# gpt-j-quantize
+
+set(TEST_TARGET gpt-j-quantize)
+add_executable(${TEST_TARGET} quantize.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml ggml_utils)

examples/gpt-j/convert-h5-to-ggml.py

@@ -47,8 +47,10 @@ def bytes_to_unicode():
     cs = [chr(n) for n in cs]
     return dict(zip(bs, cs))
 
-if len(sys.argv) < 2:
+if len(sys.argv) < 3:
     print("Usage: convert-h5-to-ggml.py dir-model [use-f32]\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
     sys.exit(1)
 
 # output in the same directory as the model
@@ -64,11 +66,21 @@ with open(dir_model + "/added_tokens.json", "r") as f:
 with open(dir_model + "/config.json", "r") as f:
     hparams = json.load(f)
 
-# use 16-bit or 32-bit floats
-use_f16 = True
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
 if len(sys.argv) > 2:
-    use_f16 = False
-    fname_out = sys.argv[1] + "/ggml-model-f32.bin"
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+    fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
+
 
 model = GPTJForCausalLM.from_pretrained(dir_model, low_cpu_mem_usage=True)
 #print (model)
@@ -85,7 +97,7 @@ fout.write(struct.pack("i", hparams["n_embd"]))
 fout.write(struct.pack("i", hparams["n_head"]))
 fout.write(struct.pack("i", hparams["n_layer"]))
 fout.write(struct.pack("i", hparams["rotary_dim"]))
-fout.write(struct.pack("i", use_f16))
+fout.write(struct.pack("i", ftype))
 
 byte_encoder = bytes_to_unicode()
 byte_decoder = {v:k for k, v in byte_encoder.items()}
@@ -114,34 +126,40 @@ for name in list_vars.keys():
     n_dims = len(data.shape);
 
     # ftype == 0 -> float32, ftype == 1 -> float16
-    ftype = 0;
-    if use_f16:
+    ftype_cur = 0;
+    if ftype != 0:
         if name[-7:] == ".weight" and n_dims == 2:
             print("  Converting to float16")
             data = data.astype(np.float16)
-            ftype = 1
+            ftype_cur = 1
         else:
             print("  Converting to float32")
             data = data.astype(np.float32)
-            ftype = 0
+            ftype_cur = 0
+    else:
+        if data.dtype != np.float32:
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype_cur = 0
 
     # for efficiency - transpose these matrices:
-    #  "transformer.h.*.mlp.fc_in.weight
-    #  "transformer.h.*.attn.out_proj.weight
+    # (note - with latest ggml this is no longer more efficient, so disabling it)
+    #  "transformer.h.*.mlp.fc_in.weight"
+    #  "transformer.h.*.attn.out_proj.weight"
     #  "transformer.h.*.attn.q_proj.weight"
     #  "transformer.h.*.attn.k_proj.weight"
     #  "transformer.h.*.attn.v_proj.weight"
-    if name.endswith(".mlp.fc_in.weight")     or \
-       name.endswith(".attn.out_proj.weight") or \
-       name.endswith(".attn.q_proj.weight")   or \
-       name.endswith(".attn.k_proj.weight")   or \
-       name.endswith(".attn.v_proj.weight"):
-        print("  Transposing")
-        data = data.transpose()
+    #if name.endswith(".mlp.fc_in.weight")     or \
+    #   name.endswith(".attn.out_proj.weight") or \
+    #   name.endswith(".attn.q_proj.weight")   or \
+    #   name.endswith(".attn.k_proj.weight")   or \
+    #   name.endswith(".attn.v_proj.weight"):
+    #    print("  Transposing")
+    #    data = data.transpose()
 
     # header
     str = name.encode('utf-8')
-    fout.write(struct.pack("iii", n_dims, len(str), ftype))
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
     for i in range(n_dims):
         fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
     fout.write(str);

examples/gpt-j/main.cpp

@@ -130,9 +130,23 @@ bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab &
         }
     }
 
-    // for the big tensors, we have the option to store the data in 16-bit floats
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
     // in order to save memory and also to speed up the computation
-    const ggml_type wtype = model.hparams.f16 ? GGML_TYPE_F16 : GGML_TYPE_F32;
+    ggml_type wtype = GGML_TYPE_COUNT;
+    switch (model.hparams.f16) {
+        case 0: wtype = GGML_TYPE_F32;  break;
+        case 1: wtype = GGML_TYPE_F16;  break;
+        case 2: wtype = GGML_TYPE_Q4_0; break;
+        case 3: wtype = GGML_TYPE_Q4_1; break;
+        default:
+                {
+                    fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n",
+                            __func__, fname.c_str(), model.hparams.f16);
+                    return false;
+                }
+    }
+
+    const ggml_type wtype2 = GGML_TYPE_F32;
 
     auto & ctx = model.ctx;
 
@@ -146,31 +160,31 @@ bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab &
         const int n_ctx   = hparams.n_ctx;
         const int n_vocab = hparams.n_vocab;
 
-        ctx_size += n_embd*ggml_type_size(GGML_TYPE_F32); // ln_f_g
-        ctx_size += n_embd*ggml_type_size(GGML_TYPE_F32); // ln_f_b
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_g
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // ln_f_b
 
-        ctx_size += n_embd*n_vocab*ggml_type_size(wtype); // wte
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // wte
 
-        ctx_size += n_embd*n_vocab*ggml_type_size(wtype);         // lmh_g
-        ctx_size +=        n_vocab*ggml_type_size(GGML_TYPE_F32); // lmh_b
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype);         // lmh_g
+        ctx_size +=        n_vocab*ggml_type_sizef(GGML_TYPE_F32); // lmh_b
 
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_g
-        ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_b
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_g
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ln_1_b
 
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_size(wtype)); // c_attn_q_proj_w
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_size(wtype)); // c_attn_k_proj_w
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_size(wtype)); // c_attn_v_proj_w
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_q_proj_w
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_k_proj_w
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_v_proj_w
 
-        ctx_size += n_layer*(n_embd*n_embd*ggml_type_size(wtype)); // c_attn_proj_w
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // c_attn_proj_w
 
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_size(wtype));         // c_mlp_fc_w
-        ctx_size += n_layer*(       4*n_embd*ggml_type_size(GGML_TYPE_F32)); // c_mlp_fc_b
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_fc_w
+        ctx_size += n_layer*(       4*n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_fc_b
 
-        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_size(wtype));         // c_mlp_proj_w_trans
-        ctx_size += n_layer*(         n_embd*ggml_type_size(GGML_TYPE_F32)); // c_mlp_proj_b
+        ctx_size += n_layer*(4*n_embd*n_embd*ggml_type_sizef(wtype));         // c_mlp_proj_w_trans
+        ctx_size += n_layer*(         n_embd*ggml_type_sizef(GGML_TYPE_F32)); // c_mlp_proj_b
 
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_size(GGML_TYPE_F32); // memory_k
-        ctx_size += n_ctx*n_layer*n_embd*ggml_type_size(GGML_TYPE_F32); // memory_v
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
 
         ctx_size += (5 + 10*n_layer)*256; // object overhead
 
@@ -231,7 +245,7 @@ bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab &
 
             layer.c_attn_proj_w         = ggml_new_tensor_2d(ctx, wtype,           n_embd,   n_embd);
 
-            layer.c_mlp_fc_w            = ggml_new_tensor_2d(ctx, wtype,         4*n_embd,   n_embd);
+            layer.c_mlp_fc_w            = ggml_new_tensor_2d(ctx, wtype,           n_embd, 4*n_embd);
             layer.c_mlp_fc_b            = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_embd);
 
             layer.c_mlp_proj_w_trans    = ggml_new_tensor_2d(ctx, wtype,         4*n_embd,   n_embd);
@@ -321,9 +335,26 @@ bool gptj_model_load(const std::string & fname, gptj_model & model, gpt_vocab &
                 return false;
             }
 
-            const size_t bpe = tensor->type == GGML_TYPE_I8 ? 1 : (ftype == 0) ? sizeof(float) : sizeof(ggml_fp16_t);
+            if (0) {
+                static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ftype_str[ftype], ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
 
-            if (nelements*bpe != ggml_nbytes(tensor)) {
+            size_t bpe = 0;
+
+            switch (ftype) {
+                case 0: bpe = ggml_type_size(GGML_TYPE_F32);  break;
+                case 1: bpe = ggml_type_size(GGML_TYPE_F16);  break;
+                case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break;
+                case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break;
+                default:
+                        {
+                            fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
+                            return false;
+                        }
+            };
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
                 fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
                         __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
                 return false;
@@ -428,9 +459,9 @@ bool gptj_eval(
 
         // self-attention
         {
-            struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, ggml_transpose(ctx0, model.layers[il].c_attn_q_proj_w), cur);
-            struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, ggml_transpose(ctx0, model.layers[il].c_attn_k_proj_w), cur);
-            struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, ggml_transpose(ctx0, model.layers[il].c_attn_v_proj_w), cur);
+            struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_q_proj_w, cur);
+            struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_k_proj_w, cur);
+            struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].c_attn_v_proj_w, cur);
 
             // store key and value to memory
             if (N >= 1) {
@@ -498,7 +529,7 @@ bool gptj_eval(
 
             // projection (no bias)
             cur = ggml_mul_mat(ctx0,
-                    ggml_transpose(ctx0, model.layers[il].c_attn_proj_w),
+                    model.layers[il].c_attn_proj_w,
                     cur);
         }
 
@@ -509,7 +540,7 @@ bool gptj_eval(
         {
             // note here we pass inpSA instead of cur
             cur = ggml_mul_mat(ctx0,
-                    ggml_transpose(ctx0, model.layers[il].c_mlp_fc_w),
+                    model.layers[il].c_mlp_fc_w,
                     inpSA);
 
             cur = ggml_add(ctx0,

examples/gpt-j/quantize.cpp

@@ -0,0 +1,327 @@
+#include "ggml/ggml.h"
+
+#include "utils.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <regex>
+
+// TODO: move somewhere else
+#define QK 32
+
+// default hparams (GPT-J 6B)
+struct gptj_hparams {
+    int32_t n_vocab = 50400;
+    int32_t n_ctx   = 2048;
+    int32_t n_embd  = 4096;
+    int32_t n_head  = 16;
+    int32_t n_layer = 28;
+    int32_t n_rot   = 64;
+    int32_t f16     = 1;
+};
+
+// quantize a model
+bool gptj_model_quantize(const std::string & fname_inp, const std::string & fname_out, int itype) {
+    ggml_type type = GGML_TYPE_Q4_1;
+
+    switch (itype) {
+        case 2: type = GGML_TYPE_Q4_0; break;
+        case 3: type = GGML_TYPE_Q4_1; break;
+        default: fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype); return 1;
+    };
+
+    if (type != GGML_TYPE_Q4_0 && type != GGML_TYPE_Q4_1) {
+        fprintf(stderr, "%s: invalid quantization type %d\n", __func__, type);
+        return false;
+    }
+
+    gpt_vocab vocab;
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *) &magic, sizeof(magic));
+        if (magic != 0x67676d6c) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *) &magic, sizeof(magic));
+    }
+
+    gptj_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        finp.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        finp.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        finp.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        finp.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        finp.read((char *) &hparams.f16,     sizeof(hparams.f16));
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: f16     = %d\n", __func__, hparams.f16);
+
+        fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        fout.write((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fout.write((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fout.write((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fout.write((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fout.write((char *) &itype,           sizeof(hparams.f16));
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        finp.read ((char *) &n_vocab, sizeof(n_vocab));
+        fout.write((char *) &n_vocab, sizeof(n_vocab));
+
+        if (n_vocab != hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname_inp.c_str(), n_vocab, hparams.n_vocab);
+            return false;
+        }
+
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            finp.read ((char *) &len, sizeof(len));
+            fout.write((char *) &len, sizeof(len));
+
+            word.resize(len);
+            finp.read ((char *) word.data(), len);
+            fout.write((char *) word.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // load weights
+    {
+        size_t total_size_org = 0;
+        size_t total_size_new = 0;
+
+        std::vector<float> work;
+
+        std::vector<uint8_t>     data_u8;
+        std::vector<ggml_fp16_t> data_f16;
+        std::vector<float>       data_f32;
+
+        std::vector<int64_t> hist_all(1 << 4, 0);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ftype;
+
+            finp.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            finp.read(reinterpret_cast<char *>(&length), sizeof(length));
+            finp.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+
+            if (finp.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                finp.read (reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            finp.read (&name[0], length);
+
+            {
+                static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
+                printf("%48s - [%5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ftype_str[ftype]);
+            }
+
+            // regexes of tensor names to be quantized
+            const std::vector<std::string> k_names = {
+                ".*weight",
+            };
+
+            bool quantize = false;
+            for (const auto & s : k_names) {
+                if (std::regex_match(name, std::regex(s))) {
+                    quantize = true;
+                    break;
+                }
+            }
+
+            // quantize only 2D tensors
+            quantize &= (n_dims == 2);
+
+            if (quantize) {
+                if (ftype != 0 && ftype != 1) {
+                    fprintf(stderr, "%s: unsupported ftype %d for integer quantization\n", __func__, ftype);
+                    return false;
+                }
+
+                if (ftype == 1) {
+                    data_f16.resize(nelements);
+                    finp.read(reinterpret_cast<char *>(data_f16.data()), nelements * sizeof(ggml_fp16_t));
+                    data_f32.resize(nelements);
+                    for (int i = 0; i < nelements; ++i) {
+                        data_f32[i] = ggml_fp16_to_fp32(data_f16[i]);
+                    }
+                } else {
+                    data_f32.resize(nelements);
+                    finp.read(reinterpret_cast<char *>(data_f32.data()), nelements * sizeof(float));
+                }
+
+                ftype = itype;
+            } else {
+                const int bpe = (ftype == 0) ? sizeof(float) : sizeof(uint16_t);
+
+                data_u8.resize(nelements*bpe);
+                finp.read(reinterpret_cast<char *>(data_u8.data()), nelements * bpe);
+            }
+
+            fout.write(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fout.write(reinterpret_cast<char *>(&length), sizeof(length));
+            fout.write(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+            for (int i = 0; i < n_dims; ++i) {
+                fout.write(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+            }
+            fout.write(&name[0], length);
+
+            if (quantize) {
+                printf("quantizing .. ");
+                work.resize(nelements); // for quantization
+
+                size_t cur_size = 0;
+                std::vector<int64_t> hist_cur(1 << 4, 0);
+
+                switch (type) {
+                    case GGML_TYPE_Q4_0:
+                        {
+                            cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], QK, hist_cur.data());
+                        } break;
+                    case GGML_TYPE_Q4_1:
+                        {
+                            cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], QK, hist_cur.data());
+                        } break;
+                    default:
+                        {
+                            fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, type);
+                            return false;
+                        }
+                }
+
+                fout.write(reinterpret_cast<char *>(work.data()), cur_size);
+                total_size_new += cur_size;
+
+                printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
+                for (int i = 0; i < hist_cur.size(); ++i) {
+                    hist_all[i] += hist_cur[i];
+                }
+
+                for (int i = 0; i < hist_cur.size(); ++i) {
+                    printf("%5.3f ", hist_cur[i] / (float)nelements);
+                }
+                printf("\n");
+            } else {
+                printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
+                fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
+                total_size_new += data_u8.size();
+            }
+
+            total_size_org += nelements * sizeof(float);
+        }
+
+        printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
+        printf("%s: quant size  = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);
+
+        {
+            int64_t sum_all = 0;
+            for (int i = 0; i < hist_all.size(); ++i) {
+                sum_all += hist_all[i];
+            }
+
+            printf("%s: hist: ", __func__);
+            for (int i = 0; i < hist_all.size(); ++i) {
+                printf("%5.3f ", hist_all[i] / (float)sum_all);
+            }
+            printf("\n");
+        }
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+// usage:
+//  ./gpt-2-quantize models/gpt-2-117M/ggml-model.bin models/gpt-2-117M/ggml-model-quant.bin type
+//
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+        fprintf(stderr, "  type = 2 - q4_0\n");
+        fprintf(stderr, "  type = 3 - q4_1\n");
+        return 1;
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const int itype = atoi(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!gptj_model_quantize(fname_inp, fname_out, itype)) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    return 0;
+}

examples/llama/CMakeLists.txt

@@ -0,0 +1,13 @@
+#
+# llama
+
+set(TEST_TARGET llama)
+add_executable(${TEST_TARGET} main.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml ggml_utils)
+
+#
+# llama-quantize
+
+set(TEST_TARGET llama-quantize)
+add_executable(${TEST_TARGET} quantize.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml ggml_utils)

examples/llama/convert-pth-to-ggml.py

@@ -0,0 +1,136 @@
+# Convert a LLaMA model checkpoint to a ggml compatible file
+#
+# Load the model using Torch
+# Iterate over all variables and write them to a binary file.
+#
+# For each variable, write the following:
+#   - Number of dimensions (int)
+#   - Name length (int)
+#   - Dimensions (int[n_dims])
+#   - Name (char[name_length])
+#   - Data (float[n_dims])
+#
+# By default, the bigger matrices are converted to 16-bit floats.
+# This can be disabled by adding the "use-f32" CLI argument.
+#
+# At the start of the ggml file we write the model parameters
+# and vocabulary.
+#
+
+import sys
+import json
+import struct
+import numpy as np
+import torch
+
+from sentencepiece import SentencePieceProcessor
+
+if len(sys.argv) < 3:
+    print("Usage: convert-ckpt-to-ggml.py dir-model ftype\n")
+    print("  ftype == 0 -> float32")
+    print("  ftype == 1 -> float16")
+    sys.exit(1)
+
+# output in the same directory as the model
+dir_model = sys.argv[1]
+fname_out = sys.argv[1] + "/ggml-model.bin"
+
+fname_hparams   = sys.argv[1] + "/params.json"
+fname_model     = sys.argv[1] + "/consolidated.00.pth"
+fname_tokenizer = sys.argv[1] + "/../tokenizer.model"
+
+# possible data types
+#   ftype == 0 -> float32
+#   ftype == 1 -> float16
+#
+# map from ftype to string
+ftype_str = ["f32", "f16"]
+
+ftype = 1
+if len(sys.argv) > 2:
+    ftype = int(sys.argv[2])
+    if ftype < 0 or ftype > 1:
+        print("Invalid ftype: " + str(ftype))
+        sys.exit(1)
+    fname_out = sys.argv[1] + "/ggml-model-" + ftype_str[ftype] + ".bin"
+
+with open(fname_hparams, "r") as f:
+    hparams = json.load(f)
+
+tokenizer = SentencePieceProcessor(fname_tokenizer)
+
+hparams.update({"vocab_size": tokenizer.vocab_size()})
+
+print(hparams)
+
+model = torch.load(fname_model, map_location="cpu")
+
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["vocab_size"]))
+fout.write(struct.pack("i", hparams["dim"]))
+fout.write(struct.pack("i", hparams["multiple_of"]))
+fout.write(struct.pack("i", hparams["n_heads"]))
+fout.write(struct.pack("i", hparams["n_layers"]))
+fout.write(struct.pack("i", 64)) # rot
+fout.write(struct.pack("i", ftype))
+
+# Is this correct??
+for i in range(32000):
+    # TODO: this is probably wrong - not sure how this tokenizer works
+    text = tokenizer.decode([29889, i]).encode('utf-8')
+    # remove the first byte (it's always '.')
+    text = text[1:]
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for k, v in model.items():
+    name = k
+    shape = v.shape
+
+    # skip layers.X.attention.inner_attention.rope.freqs
+    if name[-5:] == "freqs":
+        continue
+
+    print("Processing variable: " + name + " with shape: ", shape, " and type: ", v.dtype)
+
+    #data = tf.train.load_variable(dir_model, name).squeeze()
+    data = v.numpy().squeeze()
+    n_dims = len(data.shape);
+
+    # for efficiency - transpose some matrices
+    # "model/h.*/attn/c_attn/w"
+    # "model/h.*/attn/c_proj/w"
+    # "model/h.*/mlp/c_fc/w"
+    # "model/h.*/mlp/c_proj/w"
+    #if name[-14:] == "/attn/c_attn/w" or \
+    #   name[-14:] == "/attn/c_proj/w" or \
+    #   name[-11:] == "/mlp/c_fc/w" or \
+    #   name[-13:] == "/mlp/c_proj/w":
+    #    print("  Transposing")
+    #    data = data.transpose()
+
+    dshape = data.shape
+
+    # default type is fp16
+    ftype_cur = 1
+    if ftype == 0 or n_dims == 1:
+        print("  Converting to float32")
+        data = data.astype(np.float32)
+        ftype_cur = 0
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype_cur))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", dshape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")

examples/llama/main.cpp

@@ -0,0 +1,747 @@
+#include "ggml/ggml.h"
+
+#include "utils.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+
+// default hparams (LLaMA 7B)
+struct llama_hparams {
+    int32_t n_vocab = 32000;
+    int32_t n_ctx   = 512;   // this is provided as user input?
+    int32_t n_embd  = 4096;
+    int32_t n_mult  = 256;
+    int32_t n_head  = 32;
+    int32_t n_layer = 32;
+    int32_t n_rot   = 64;
+    int32_t f16     = 1;
+};
+
+struct llama_layer {
+    // normalization
+    struct ggml_tensor * attention_norm;
+
+    // attention
+    struct ggml_tensor * wq;
+    struct ggml_tensor * wk;
+    struct ggml_tensor * wv;
+    struct ggml_tensor * wo;
+
+    // normalization
+    struct ggml_tensor * ffn_norm;
+
+    // ff
+    struct ggml_tensor * w1;
+    struct ggml_tensor * w2;
+    struct ggml_tensor * w3;
+};
+
+struct llama_model {
+    llama_hparams hparams;
+
+    struct ggml_tensor * tok_embeddings;
+
+    struct ggml_tensor * norm;
+    struct ggml_tensor * output;
+
+    std::vector<llama_layer> layers;
+
+    // key + value memory
+    struct ggml_tensor * memory_k;
+    struct ggml_tensor * memory_v;
+
+    //
+    struct ggml_context * ctx;
+    std::map<std::string, struct ggml_tensor *> tensors;
+};
+
+// load the model's weights from a file
+bool llama_model_load(const std::string & fname, llama_model & model, gpt_vocab & vocab, int n_ctx) {
+    printf("%s: loading model from '%s' - please wait ...\n", __func__, fname.c_str());
+
+    auto fin = std::ifstream(fname, std::ios::binary);
+    if (!fin) {
+        fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        fin.read((char *) &magic, sizeof(magic));
+        if (magic != 0x67676d6c) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
+            return false;
+        }
+    }
+
+    int n_ff = 0;
+
+    // load hparams
+    {
+        auto & hparams = model.hparams;
+
+        fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        //fin.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fin.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fin.read((char *) &hparams.n_mult,  sizeof(hparams.n_mult));
+        fin.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fin.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fin.read((char *) &hparams.f16,     sizeof(hparams.f16));
+
+        hparams.n_ctx = n_ctx;
+
+        n_ff = ((2*(4*hparams.n_embd)/3 + hparams.n_mult - 1)/hparams.n_mult)*hparams.n_mult;
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_mult  = %d\n", __func__, hparams.n_mult);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: n_rot   = %d\n", __func__, hparams.n_rot);
+        printf("%s: f16     = %d\n", __func__, hparams.f16);
+        printf("%s: n_ff    = %d\n", __func__, n_ff);
+    }
+
+    // load vocab
+    {
+        const int32_t n_vocab = model.hparams.n_vocab;
+
+        if (n_vocab != model.hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
+            return false;
+        }
+
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            fin.read((char *) &len, sizeof(len));
+
+            word.resize(len);
+            fin.read((char *) word.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+
+            //if (i < 30000) {
+            //    printf("%s: vocab[%d] = '%s'\n", __func__, i, word.c_str());
+            //}
+        }
+    }
+
+    // for the big tensors, we have the option to store the data in 16-bit floats or quantized
+    // in order to save memory and also to speed up the computation
+    ggml_type wtype = GGML_TYPE_COUNT;
+    switch (model.hparams.f16) {
+        case 0: wtype = GGML_TYPE_F32;  break;
+        case 1: wtype = GGML_TYPE_F16;  break;
+        case 2: wtype = GGML_TYPE_Q4_0; break;
+        case 3: wtype = GGML_TYPE_Q4_1; break;
+        default:
+                {
+                    fprintf(stderr, "%s: invalid model file '%s' (bad f16 value %d)\n",
+                            __func__, fname.c_str(), model.hparams.f16);
+                    return false;
+                }
+    }
+
+    const ggml_type wtype2 = GGML_TYPE_F32;
+
+    auto & ctx = model.ctx;
+
+    size_t ctx_size = 0;
+
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // tok_embeddings
+
+        ctx_size += n_embd*ggml_type_sizef(GGML_TYPE_F32); // norm
+
+        ctx_size += n_embd*n_vocab*ggml_type_sizef(wtype); // output
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // attention_norm
+
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wq
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wk
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wv
+        ctx_size += n_layer*(n_embd*n_embd*ggml_type_sizef(wtype)); // wo
+
+        ctx_size += n_layer*(n_embd*ggml_type_sizef(GGML_TYPE_F32)); // ffn_norm
+
+        ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w1
+        ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w2
+        ctx_size += n_layer*(n_ff*n_embd*ggml_type_sizef(wtype)); // w3
+
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_k
+        ctx_size += n_ctx*n_layer*n_embd*ggml_type_sizef(GGML_TYPE_F32); // memory_v
+
+        ctx_size += (5 + 10*n_layer)*256; // object overhead
+
+        printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
+    }
+
+    // create the ggml context
+    {
+        struct ggml_init_params params = {
+            .mem_size   = ctx_size,
+            .mem_buffer = NULL,
+        };
+
+        model.ctx = ggml_init(params);
+        if (!model.ctx) {
+            fprintf(stderr, "%s: ggml_init() failed\n", __func__);
+            return false;
+        }
+    }
+
+    // prepare memory for the weights
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+        const int n_vocab = hparams.n_vocab;
+
+        model.layers.resize(n_layer);
+
+        model.tok_embeddings = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+
+        model.norm   = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
+        model.output = ggml_new_tensor_2d(ctx, wtype,         n_embd, n_vocab);
+
+        // map by name
+        model.tensors["tok_embeddings.weight"] = model.tok_embeddings;
+
+        model.tensors["norm.weight"]   = model.norm;
+        model.tensors["output.weight"] = model.output;
+
+        for (int i = 0; i < n_layer; ++i) {
+            auto & layer = model.layers[i];
+
+            layer.attention_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.wq = ggml_new_tensor_2d(ctx, wtype, n_embd,   n_embd);
+            layer.wk = ggml_new_tensor_2d(ctx, wtype, n_embd,   n_embd);
+            layer.wv = ggml_new_tensor_2d(ctx, wtype, n_embd,   n_embd);
+            layer.wo = ggml_new_tensor_2d(ctx, wtype, n_embd,   n_embd);
+
+            layer.ffn_norm = ggml_new_tensor_1d(ctx, GGML_TYPE_F32,   n_embd);
+
+            layer.w1 = ggml_new_tensor_2d(ctx, wtype, n_embd,   n_ff);
+            layer.w2 = ggml_new_tensor_2d(ctx, wtype,   n_ff, n_embd);
+            layer.w3 = ggml_new_tensor_2d(ctx, wtype, n_embd,   n_ff);
+
+            // map by name
+            model.tensors["layers." + std::to_string(i) + ".attention_norm.weight"] = layer.attention_norm;
+
+            model.tensors["layers." + std::to_string(i) + ".attention.wq.weight"] = layer.wq;
+            model.tensors["layers." + std::to_string(i) + ".attention.wk.weight"] = layer.wk;
+            model.tensors["layers." + std::to_string(i) + ".attention.wv.weight"] = layer.wv;
+            model.tensors["layers." + std::to_string(i) + ".attention.wo.weight"] = layer.wo;
+
+            model.tensors["layers." + std::to_string(i) + ".ffn_norm.weight"] = layer.ffn_norm;
+
+            model.tensors["layers." + std::to_string(i) + ".feed_forward.w1.weight"] = layer.w1;
+            model.tensors["layers." + std::to_string(i) + ".feed_forward.w2.weight"] = layer.w2;
+            model.tensors["layers." + std::to_string(i) + ".feed_forward.w3.weight"] = layer.w3;
+        }
+    }
+
+    // key + value memory
+    {
+        const auto & hparams = model.hparams;
+
+        const int n_embd  = hparams.n_embd;
+        const int n_layer = hparams.n_layer;
+        const int n_ctx   = hparams.n_ctx;
+
+        const int n_mem      = n_layer*n_ctx;
+        const int n_elements = n_embd*n_mem;
+
+        model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
+        model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
+
+        const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
+
+        printf("%s: memory_size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
+    }
+
+    // load weights
+    {
+        int n_tensors = 0;
+        size_t total_size = 0;
+
+        printf("%s: ", __func__);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ftype;
+
+            fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fin.read(reinterpret_cast<char *>(&length), sizeof(length));
+            fin.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+
+            if (fin.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            fin.read(&name[0], length);
+
+            if (model.tensors.find(name.data()) == model.tensors.end()) {
+                fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
+                return false;
+            }
+
+            auto tensor = model.tensors[name.data()];
+            if (ggml_nelements(tensor) != nelements) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
+                return false;
+            }
+
+            if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
+                fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
+                        __func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
+                return false;
+            }
+
+            if (0) {
+                static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
+                printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ftype_str[ftype], ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
+            }
+
+            size_t bpe = 0;
+
+            switch (ftype) {
+                case 0: bpe = ggml_type_size(GGML_TYPE_F32);  break;
+                case 1: bpe = ggml_type_size(GGML_TYPE_F16);  break;
+                case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break;
+                case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break;
+                default:
+                        {
+                            fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
+                            return false;
+                        }
+            };
+
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
+                fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
+                        __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
+                return false;
+            }
+
+            fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
+
+            //printf("%42s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
+            total_size += ggml_nbytes(tensor);
+            if (++n_tensors % 8 == 0) {
+                printf(".");
+                fflush(stdout);
+            }
+        }
+
+        printf(" done\n");
+
+        printf("%s: model size = %8.2f MB / num tensors = %d\n", __func__, total_size/1024.0/1024.0, n_tensors);
+    }
+
+    fin.close();
+
+    return true;
+}
+
+// evaluate the transformer
+//
+//   - model:     the model
+//   - n_threads: number of threads to use
+//   - n_past:    the context size so far
+//   - embd_inp:  the embeddings of the tokens in the context
+//   - embd_w:    the predicted logits for the next token
+//
+// The GPT-J model requires about 16MB of memory per input token.
+//
+bool llama_eval(
+        const llama_model & model,
+        const int n_threads,
+        const int n_past,
+        const std::vector<gpt_vocab::id> & embd_inp,
+              std::vector<float>         & embd_w,
+              size_t                     & mem_per_token) {
+    const int N = embd_inp.size();
+
+    const auto & hparams = model.hparams;
+
+    const int n_embd  = hparams.n_embd;
+    const int n_layer = hparams.n_layer;
+    const int n_ctx   = hparams.n_ctx;
+    const int n_head  = hparams.n_head;
+    const int n_vocab = hparams.n_vocab;
+    const int n_rot   = hparams.n_rot;
+
+    const int d_key = n_embd/n_head;
+
+    static size_t buf_size = 256u*1024*1024;
+    static void * buf = malloc(buf_size);
+
+    if (mem_per_token > 0 && mem_per_token*N > buf_size) {
+        const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
+        //printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
+
+        // reallocate
+        buf_size = buf_size_new;
+        buf = realloc(buf, buf_size);
+        if (buf == nullptr) {
+            fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
+            return false;
+        }
+    }
+
+    struct ggml_init_params params = {
+        .mem_size   = buf_size,
+        .mem_buffer = buf,
+    };
+
+    struct ggml_context * ctx0 = ggml_init(params);
+    struct ggml_cgraph gf = { .n_threads = n_threads };
+
+    struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
+    memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
+
+    struct ggml_tensor * inpL = ggml_get_rows(ctx0, model.tok_embeddings, embd);
+
+    for (int il = 0; il < n_layer; ++il) {
+        struct ggml_tensor * inpSA = inpL;
+
+        struct ggml_tensor * cur;
+
+        // norm
+        {
+            cur = ggml_norm(ctx0, inpL);
+
+            // cur = attention_norm*cur
+            cur = ggml_mul(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].attention_norm, cur),
+                        cur);
+        }
+
+        // self-attention
+        {
+            struct ggml_tensor * Qcur = ggml_mul_mat(ctx0, model.layers[il].wq, cur);
+            struct ggml_tensor * Kcur = ggml_mul_mat(ctx0, model.layers[il].wk, cur);
+            struct ggml_tensor * Vcur = ggml_mul_mat(ctx0, model.layers[il].wv, cur);
+
+            // store key and value to memory
+            if (N >= 1) {
+                struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
+                struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
+
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
+                ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
+            }
+
+            // Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
+            struct ggml_tensor * Q =
+                ggml_permute(ctx0,
+                        ggml_rope(ctx0,
+                            ggml_cpy(ctx0,
+                                Qcur,
+                                ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
+                            n_past, n_rot, 0),
+                        0, 2, 1, 3);
+
+            // K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
+            struct ggml_tensor * K =
+                ggml_permute(ctx0,
+                        ggml_rope(ctx0,
+                            ggml_reshape_3d(ctx0,
+                                ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
+                                n_embd/n_head, n_head, n_past + N),
+                            n_past, n_rot, 1),
+                        0, 2, 1, 3);
+
+            // K * Q
+            struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
+
+            // KQ_scaled = KQ / sqrt(n_embd/n_head)
+            struct ggml_tensor * KQ_scaled =
+                ggml_scale(ctx0,
+                        KQ,
+                        ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
+                        );
+
+            // KQ_masked = mask_past(KQ_scaled)
+            struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
+
+            // KQ = soft_max(KQ_masked)
+            struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
+
+            // V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
+            struct ggml_tensor * V_trans =
+                ggml_permute(ctx0,
+                        ggml_reshape_3d(ctx0,
+                            ggml_view_1d(ctx0, model.memory_v, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_v)*n_embd),
+                            n_embd/n_head, n_head, n_past + N),
+                        1, 2, 0, 3);
+
+            // KQV = transpose(V) * KQ_soft_max
+            struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
+
+            // KQV_merged = KQV.permute(0, 2, 1, 3)
+            struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
+
+            // cur = KQV_merged.contiguous().view(n_embd, N)
+            cur = ggml_cpy(ctx0,
+                    KQV_merged,
+                    ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
+
+            // projection (no bias)
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].wo,
+                    cur);
+        }
+
+        struct ggml_tensor * inpFF = ggml_add(ctx0, cur, inpSA);
+
+        // feed-forward network
+        {
+            // norm
+            {
+                cur = ggml_norm(ctx0, inpFF);
+
+                // cur = ffn_norm*cur
+                cur = ggml_mul(ctx0,
+                        ggml_repeat(ctx0, model.layers[il].ffn_norm, cur),
+                        cur);
+            }
+
+            struct ggml_tensor * tmp = ggml_mul_mat(ctx0,
+                    model.layers[il].w3,
+                    cur);
+
+
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].w1,
+                    cur);
+
+            // SILU activation
+            cur = ggml_silu(ctx0, cur);
+
+            cur = ggml_mul(ctx0, cur, tmp);
+
+            cur = ggml_mul_mat(ctx0,
+                    model.layers[il].w2,
+                    cur);
+        }
+
+        cur  = ggml_add(ctx0, cur, inpFF);
+
+        // input for next layer
+        inpL = cur;
+    }
+
+    // norm
+    {
+        inpL = ggml_norm(ctx0, inpL);
+
+        // inpL = norm*inpL
+        inpL = ggml_mul(ctx0,
+                    ggml_repeat(ctx0, model.norm, inpL),
+                    inpL);
+    }
+
+    // lm_head
+    {
+        inpL = ggml_mul_mat(ctx0, model.output, inpL);
+    }
+
+    // logits -> probs
+    //inpL = ggml_soft_max(ctx0, inpL);
+
+    // run the computation
+    ggml_build_forward_expand(&gf, inpL);
+    ggml_graph_compute       (ctx0, &gf);
+
+    //if (n_past%100 == 0) {
+    //    ggml_graph_print   (&gf);
+    //    ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
+    //}
+
+    //embd_w.resize(n_vocab*N);
+    //memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
+
+    // return result for just the last token
+    embd_w.resize(n_vocab);
+    memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
+
+    if (mem_per_token == 0) {
+        mem_per_token = ggml_used_mem(ctx0)/N;
+    }
+    //printf("used_mem = %zu\n", ggml_used_mem(ctx0));
+
+    ggml_free(ctx0);
+
+    return true;
+}
+
+int main(int argc, char ** argv) {
+    const int64_t t_main_start_us = ggml_time_us();
+
+    gpt_params params;
+    params.model = "models/llama-7B/ggml-model.bin";
+
+    if (gpt_params_parse(argc, argv, params) == false) {
+        return 1;
+    }
+
+    if (params.seed < 0) {
+        params.seed = time(NULL);
+    }
+
+    printf("%s: seed = %d\n", __func__, params.seed);
+
+    std::mt19937 rng(params.seed);
+    if (params.prompt.empty()) {
+        params.prompt = gpt_random_prompt(rng);
+    }
+
+    int64_t t_load_us = 0;
+
+    gpt_vocab vocab;
+    llama_model model;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!llama_model_load(params.model, model, vocab, 512)) {  // TODO: set context from user input ??
+            fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
+            return 1;
+        }
+
+        t_load_us = ggml_time_us() - t_start_us;
+    }
+
+    int n_past = 0;
+
+    int64_t t_sample_us  = 0;
+    int64_t t_predict_us = 0;
+
+    std::vector<float> logits;
+
+    // tokenize the prompt
+    std::vector<gpt_vocab::id> embd_inp = ::llama_tokenize(vocab, params.prompt, true);
+
+    params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
+
+    printf("%s: prompt: '%s'\n", __func__, params.prompt.c_str());
+    printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
+    for (int i = 0; i < (int) embd_inp.size(); i++) {
+        printf("%6d -> '%s'\n", embd_inp[i], vocab.id_to_token.at(embd_inp[i]).c_str());
+    }
+    printf("\n\n");
+
+    std::vector<gpt_vocab::id> embd;
+
+    // determine the required inference memory per token:
+    size_t mem_per_token = 0;
+    llama_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, logits, mem_per_token);
+
+    for (int i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
+        // predict
+        if (embd.size() > 0) {
+            const int64_t t_start_us = ggml_time_us();
+
+            if (!llama_eval(model, params.n_threads, n_past, embd, logits, mem_per_token)) {
+                printf("Failed to predict\n");
+                return 1;
+            }
+
+            t_predict_us += ggml_time_us() - t_start_us;
+        }
+
+        n_past += embd.size();
+        embd.clear();
+
+        if (i >= embd_inp.size()) {
+            // sample next token
+            const int   top_k = params.top_k;
+            const float top_p = params.top_p;
+            const float temp  = params.temp;
+
+            const int n_vocab = model.hparams.n_vocab;
+
+            gpt_vocab::id id = 0;
+
+            {
+                const int64_t t_start_sample_us = ggml_time_us();
+
+                id = gpt_sample_top_k_top_p(vocab, logits.data() + (logits.size() - n_vocab), top_k, top_p, temp, rng);
+
+                t_sample_us += ggml_time_us() - t_start_sample_us;
+            }
+
+            // add it to the context
+            embd.push_back(id);
+        } else {
+            // if here, it means we are still processing the input prompt
+            for (int k = i; k < embd_inp.size(); k++) {
+                embd.push_back(embd_inp[k]);
+                if (embd.size() > params.n_batch) {
+                    break;
+                }
+            }
+            i += embd.size() - 1;
+        }
+
+        // display text
+        for (auto id : embd) {
+            printf("%s", vocab.id_to_token[id].c_str());
+        }
+        fflush(stdout);
+
+        // end of text token
+        if (embd.back() == 50256) {
+            break;
+        }
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n\n");
+        printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
+        printf("%s:     load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
+        printf("%s:   sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
+        printf("%s:  predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    ggml_free(model.ctx);
+
+    return 0;
+}

examples/llama/quantize.cpp

@@ -0,0 +1,330 @@
+#include "ggml/ggml.h"
+
+#include "utils.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <regex>
+
+// TODO: move somewhere else
+#define QK 32
+
+// default hparams (LLaMA76B)
+struct llama_hparams {
+    int32_t n_vocab = 32000;
+    int32_t n_ctx   = 512;   // this is provided as user input?
+    int32_t n_embd  = 4096;
+    int32_t n_mult  = 256;
+    int32_t n_head  = 32;
+    int32_t n_layer = 32;
+    int32_t n_rot   = 64;
+    int32_t f16     = 1;
+};
+
+
+// quantize a model
+bool llama_model_quantize(const std::string & fname_inp, const std::string & fname_out, int itype) {
+    ggml_type type = GGML_TYPE_Q4_1;
+
+    switch (itype) {
+        case 2: type = GGML_TYPE_Q4_0; break;
+        case 3: type = GGML_TYPE_Q4_1; break;
+        default: fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype); return 1;
+    };
+
+    if (type != GGML_TYPE_Q4_0 && type != GGML_TYPE_Q4_1) {
+        fprintf(stderr, "%s: invalid quantization type %d\n", __func__, type);
+        return false;
+    }
+
+    gpt_vocab vocab;
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *) &magic, sizeof(magic));
+        if (magic != 0x67676d6c) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *) &magic, sizeof(magic));
+    }
+
+    llama_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        //finp.read((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        finp.read((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        finp.read((char *) &hparams.n_mult,  sizeof(hparams.n_mult));
+        finp.read((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        finp.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        finp.read((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        finp.read((char *) &hparams.f16,     sizeof(hparams.f16));
+
+        printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
+        printf("%s: n_ctx   = %d\n", __func__, hparams.n_ctx);
+        printf("%s: n_embd  = %d\n", __func__, hparams.n_embd);
+        printf("%s: n_mult  = %d\n", __func__, hparams.n_mult);
+        printf("%s: n_head  = %d\n", __func__, hparams.n_head);
+        printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
+        printf("%s: f16     = %d\n", __func__, hparams.f16);
+
+        fout.write((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
+        //fout.write((char *) &hparams.n_ctx,   sizeof(hparams.n_ctx));
+        fout.write((char *) &hparams.n_embd,  sizeof(hparams.n_embd));
+        fout.write((char *) &hparams.n_mult,  sizeof(hparams.n_mult));
+        fout.write((char *) &hparams.n_head,  sizeof(hparams.n_head));
+        fout.write((char *) &hparams.n_layer, sizeof(hparams.n_layer));
+        fout.write((char *) &hparams.n_rot,   sizeof(hparams.n_rot));
+        fout.write((char *) &itype,           sizeof(hparams.f16));
+    }
+
+    // load vocab
+    {
+        const int32_t n_vocab = hparams.n_vocab;
+
+        if (n_vocab != hparams.n_vocab) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+                    __func__, fname_inp.c_str(), n_vocab, hparams.n_vocab);
+            return false;
+        }
+
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            finp.read ((char *) &len, sizeof(len));
+            fout.write((char *) &len, sizeof(len));
+
+            word.resize(len);
+            finp.read ((char *) word.data(), len);
+            fout.write((char *) word.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // load weights
+    {
+        size_t total_size_org = 0;
+        size_t total_size_new = 0;
+
+        std::vector<float> work;
+
+        std::vector<uint8_t>     data_u8;
+        std::vector<ggml_fp16_t> data_f16;
+        std::vector<float>       data_f32;
+
+        std::vector<int64_t> hist_all(1 << 4, 0);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ftype;
+
+            finp.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            finp.read(reinterpret_cast<char *>(&length), sizeof(length));
+            finp.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+
+            if (finp.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[2] = { 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                finp.read (reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            finp.read (&name[0], length);
+
+            {
+                static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
+                printf("%48s - [%5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ftype_str[ftype]);
+            }
+
+            // regexes of tensor names to be quantized
+            const std::vector<std::string> k_names = {
+                ".*weight",
+            };
+
+            bool quantize = false;
+            for (const auto & s : k_names) {
+                if (std::regex_match(name, std::regex(s))) {
+                    quantize = true;
+                    break;
+                }
+            }
+
+            // quantize only 2D tensors
+            quantize &= (n_dims == 2);
+
+            if (quantize) {
+                if (ftype != 0 && ftype != 1) {
+                    fprintf(stderr, "%s: unsupported ftype %d for integer quantization\n", __func__, ftype);
+                    return false;
+                }
+
+                if (ftype == 1) {
+                    data_f16.resize(nelements);
+                    finp.read(reinterpret_cast<char *>(data_f16.data()), nelements * sizeof(ggml_fp16_t));
+                    data_f32.resize(nelements);
+                    for (int i = 0; i < nelements; ++i) {
+                        data_f32[i] = ggml_fp16_to_fp32(data_f16[i]);
+                    }
+                } else {
+                    data_f32.resize(nelements);
+                    finp.read(reinterpret_cast<char *>(data_f32.data()), nelements * sizeof(float));
+                }
+
+                ftype = itype;
+            } else {
+                const int bpe = (ftype == 0) ? sizeof(float) : sizeof(uint16_t);
+
+                data_u8.resize(nelements*bpe);
+                finp.read(reinterpret_cast<char *>(data_u8.data()), nelements * bpe);
+            }
+
+            fout.write(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fout.write(reinterpret_cast<char *>(&length), sizeof(length));
+            fout.write(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+            for (int i = 0; i < n_dims; ++i) {
+                fout.write(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+            }
+            fout.write(&name[0], length);
+
+            if (quantize) {
+                printf("quantizing .. ");
+                work.resize(nelements); // for quantization
+
+                size_t cur_size = 0;
+                std::vector<int64_t> hist_cur(1 << 4, 0);
+
+                switch (type) {
+                    case GGML_TYPE_Q4_0:
+                        {
+                            cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], QK, hist_cur.data());
+                        } break;
+                    case GGML_TYPE_Q4_1:
+                        {
+                            cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], QK, hist_cur.data());
+                        } break;
+                    default:
+                        {
+                            fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, type);
+                            return false;
+                        }
+                }
+
+                fout.write(reinterpret_cast<char *>(work.data()), cur_size);
+                total_size_new += cur_size;
+
+                printf("size = %8.2f MB -> %8.2f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
+                for (int i = 0; i < hist_cur.size(); ++i) {
+                    hist_all[i] += hist_cur[i];
+                }
+
+                for (int i = 0; i < hist_cur.size(); ++i) {
+                    printf("%5.3f ", hist_cur[i] / (float)nelements);
+                }
+                printf("\n");
+            } else {
+                printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
+                fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
+                total_size_new += data_u8.size();
+            }
+
+            total_size_org += nelements * sizeof(float);
+        }
+
+        printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
+        printf("%s: quant size  = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);
+
+        {
+            int64_t sum_all = 0;
+            for (int i = 0; i < hist_all.size(); ++i) {
+                sum_all += hist_all[i];
+            }
+
+            printf("%s: hist: ", __func__);
+            for (int i = 0; i < hist_all.size(); ++i) {
+                printf("%5.3f ", hist_all[i] / (float)sum_all);
+            }
+            printf("\n");
+        }
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+// usage:
+//  ./llama-quantize models/llama/ggml-model.bin models/llama/ggml-model-quant.bin type
+//
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+        fprintf(stderr, "  type = 2 - q4_0\n");
+        fprintf(stderr, "  type = 3 - q4_1\n");
+        return 1;
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const int itype = atoi(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!llama_model_quantize(fname_inp, fname_out, itype)) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    return 0;
+}

examples/utils.cpp

@@ -230,6 +230,38 @@ std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::stri
     return tokens;
 }
 
+std::vector<gpt_vocab::id> llama_tokenize(const gpt_vocab & vocab, const std::string & text, bool bos) {
+    std::vector<gpt_vocab::id> res;
+
+    if (bos) {
+        res.push_back(1); // TODO: replace with vocab.bos
+    }
+
+    // find the longest token that matches the text
+    int pos = 0;
+    while (true) {
+        int l = 0;
+        int t = 0;
+        for (const auto & kv : vocab.id_to_token) {
+            if (kv.second.size() < l) continue;
+            if (kv.second.size() > text.size() - pos) continue;
+            if (text.substr(pos, kv.second.size()) == kv.second) {
+                l = kv.second.size();
+                t = kv.first;
+            }
+        }
+
+        if (l == 0) {
+            break;
+        }
+
+        res.push_back(t);
+        pos += l;
+    }
+
+    return res;
+}
+
 bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab) {
     printf("%s: loading vocab from '%s'\n", __func__, fname.c_str());
 
@@ -328,3 +360,113 @@ gpt_vocab::id gpt_sample_top_k_top_p(
 
     return logits_id[idx].second;
 }
+
+size_t ggml_quantize_q4_0(float * src, void * dst, int n, int k, int qk, int64_t * hist) {
+    const int nb = k / qk;
+    const size_t row_size = nb*(sizeof(float) + sizeof(uint8_t)*qk/2);
+
+    assert(k % qk == 0);
+
+    uint8_t pp[qk/2];
+
+    char * pdst = (char *) dst;
+
+    for (int j = 0; j < n; j += k) {
+        float   * pd = (float *)   (pdst + (j/k)*row_size);
+        uint8_t * pb = (uint8_t *) (pd + nb);
+
+        for (int i = 0; i < nb; i++) {
+            float amax = 0.0f; // absolute max
+
+            {
+                for (int l = 0; l < qk; l++) {
+                    const float v = src[j + i*qk + l];
+                    amax = std::max(amax, fabsf(v));
+                }
+
+                const float d = amax / ((1 << 3) - 1);
+                const float id = d ? 1.0f/d : 0.0f;
+
+                pd[i] = d;
+
+                for (int l = 0; l < qk; l += 2) {
+                    const float v0 = (src[j + i*qk + l + 0])*id;
+                    const float v1 = (src[j + i*qk + l + 1])*id;
+
+                    const uint8_t vi0 = ((int8_t) (round(v0))) + 8;
+                    const uint8_t vi1 = ((int8_t) (round(v1))) + 8;
+
+                    assert(vi0 >= 0 && vi0 < 16);
+                    assert(vi1 >= 0 && vi1 < 16);
+
+                    hist[vi0]++;
+                    hist[vi1]++;
+
+                    pp[l/2] = vi0 | (vi1 << 4);
+                }
+
+                memcpy(pb + i*qk/2, pp, sizeof(pp));
+            }
+        }
+    }
+
+    return (n/k)*row_size;
+}
+
+size_t ggml_quantize_q4_1(float * src, void * dst, int n, int k, int qk, int64_t * hist) {
+    const int nb = k / qk;
+    const size_t row_size = nb*(2*sizeof(float) + sizeof(uint8_t)*qk/2);
+
+    assert(k % qk == 0);
+
+    uint8_t pp[qk/2];
+
+    char * pdst = (char *) dst;
+
+    for (int j = 0; j < n; j += k) {
+        float   * pm = (float *)   (pdst + (j/k)*row_size);
+        float   * pd = (float *)   (pm + nb);
+        uint8_t * pb = (uint8_t *) (pd + nb);
+
+        //printf("n = %d, k = %d, nb = %d, row_size = %d, j = %d, pm = %p, pd = %p, pb = %p\n", n, k, nb, row_size, j, pm, pd, pb);
+
+        for (int i = 0; i < nb; i++) {
+            float min = std::numeric_limits<float>::max();
+            float max = std::numeric_limits<float>::min();
+
+            {
+                for (int l = 0; l < qk; l++) {
+                    const float v = src[j + i*qk + 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;
+
+                pm[i] = min;
+                pd[i] = d;
+
+                for (int l = 0; l < qk; l += 2) {
+                    const float v0 = (src[j + i*qk + l + 0] - min)*id;
+                    const float v1 = (src[j + i*qk + l + 1] - min)*id;
+
+                    const uint8_t vi0 = round(v0);
+                    const uint8_t vi1 = round(v1);
+
+                    assert(vi0 >= 0 && vi0 < 16);
+                    assert(vi1 >= 0 && vi1 < 16);
+
+                    hist[vi0]++;
+                    hist[vi1]++;
+
+                    pp[l/2] = vi0 | (vi1 << 4);
+                }
+
+                memcpy(pb + i*qk/2, pp, sizeof(pp));
+            }
+        }
+    }
+
+    return (n/k)*row_size;
+}

examples/utils.h

@@ -63,6 +63,10 @@ std::map<std::string, int32_t> json_parse(const std::string & fname);
 //
 std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text);
 
+// TODO: this is probably wrong, but I cannot figure out how this tokenizer works ..
+// ref: https://github.com/google/sentencepiece
+std::vector<gpt_vocab::id> llama_tokenize(const gpt_vocab & vocab, const std::string & text, bool bos);
+
 // load the tokens from encoder.json
 bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab);
 
@@ -82,3 +86,9 @@ gpt_vocab::id gpt_sample_top_k_top_p(
         double temp,
         std::mt19937 & rng);
 
+//
+// Quantization
+//
+
+size_t ggml_quantize_q4_0(float * src, void * dst, int n, int k, int qk, int64_t * hist);
+size_t ggml_quantize_q4_1(float * src, void * dst, int n, int k, int qk, int64_t * hist);

examples/whisper/CMakeLists.txt

@@ -13,3 +13,10 @@ set(TEST_TARGET whisper)
 add_executable(${TEST_TARGET} main.cpp common.cpp)
 target_link_libraries(${TEST_TARGET} PRIVATE whisper-cpp)
 target_include_directories(${TEST_TARGET} PRIVATE ${CMAKE_CURRENT_SOURCE_DIR}/..)
+
+#
+# whisper-quantize
+
+set(TEST_TARGET whisper-quantize)
+add_executable(${TEST_TARGET} quantize.cpp)
+target_link_libraries(${TEST_TARGET} PRIVATE ggml ggml_utils)

examples/whisper/convert-pt-to-ggml.py

@@ -303,8 +303,9 @@ for name in list_vars.keys():
             data = data.astype(np.float32)
             ftype = 0
     else:
-        data = data.astype(np.float32)
-        ftype = 0
+        if n_dims < 3 and data.dtype != np.float32:
+            data = data.astype(np.float32)
+            ftype = 0
 
     #if name.startswith("encoder"):
     #    if name.endswith("mlp.0.weight") or \

examples/whisper/quantize.cpp

@@ -0,0 +1,376 @@
+#include "ggml/ggml.h"
+
+#include "utils.h"
+
+#include <cassert>
+#include <cmath>
+#include <cstdio>
+#include <cstring>
+#include <fstream>
+#include <map>
+#include <string>
+#include <vector>
+#include <regex>
+
+// TODO: move somewhere else
+#define QK 32
+
+// default hparams (Whisper tiny)
+struct whisper_hparams {
+    int32_t n_vocab       = 51864;
+    int32_t n_audio_ctx   = 1500;
+    int32_t n_audio_state = 384;
+    int32_t n_audio_head  = 6;
+    int32_t n_audio_layer = 4;
+    int32_t n_text_ctx    = 448;
+    int32_t n_text_state  = 384;
+    int32_t n_text_head   = 6;
+    int32_t n_text_layer  = 4;
+    int32_t n_mels        = 80;
+    int32_t f16           = 1;
+};
+
+struct whisper_filters {
+    int32_t n_mel;
+    int32_t n_fft;
+
+    std::vector<float> data;
+};
+
+// quantize a model
+bool whisper_model_quantize(const std::string & fname_inp, const std::string & fname_out, int itype) {
+    ggml_type type = GGML_TYPE_Q4_1;
+
+    switch (itype) {
+        case 2: type = GGML_TYPE_Q4_0; break;
+        case 3: type = GGML_TYPE_Q4_1; break;
+        default: fprintf(stderr, "%s: invalid quantization type %d\n", __func__, itype); return 1;
+    };
+
+    if (type != GGML_TYPE_Q4_0 && type != GGML_TYPE_Q4_1) {
+        fprintf(stderr, "%s: invalid quantization type %d\n", __func__, type);
+        return false;
+    }
+
+    gpt_vocab vocab;
+
+    printf("%s: loading model from '%s'\n", __func__, fname_inp.c_str());
+
+    auto finp = std::ifstream(fname_inp, std::ios::binary);
+    if (!finp) {
+        fprintf(stderr, "%s: failed to open '%s' for reading\n", __func__, fname_inp.c_str());
+        return false;
+    }
+
+    auto fout = std::ofstream(fname_out, std::ios::binary);
+    if (!fout) {
+        fprintf(stderr, "%s: failed to open '%s' for writing\n", __func__, fname_out.c_str());
+        return false;
+    }
+
+    // verify magic
+    {
+        uint32_t magic;
+        finp.read((char *) &magic, sizeof(magic));
+        if (magic != 0x67676d6c) {
+            fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname_inp.c_str());
+            return false;
+        }
+
+        fout.write((char *) &magic, sizeof(magic));
+    }
+
+    whisper_hparams hparams;
+
+    // load hparams
+    {
+        finp.read((char *) &hparams.n_vocab,       sizeof(hparams.n_vocab));
+        finp.read((char *) &hparams.n_audio_ctx,   sizeof(hparams.n_audio_ctx));
+        finp.read((char *) &hparams.n_audio_state, sizeof(hparams.n_audio_state));
+        finp.read((char *) &hparams.n_audio_head,  sizeof(hparams.n_audio_head));
+        finp.read((char *) &hparams.n_audio_layer, sizeof(hparams.n_audio_layer));
+        finp.read((char *) &hparams.n_text_ctx,    sizeof(hparams.n_text_ctx));
+        finp.read((char *) &hparams.n_text_state,  sizeof(hparams.n_text_state));
+        finp.read((char *) &hparams.n_text_head,   sizeof(hparams.n_text_head));
+        finp.read((char *) &hparams.n_text_layer,  sizeof(hparams.n_text_layer));
+        finp.read((char *) &hparams.n_mels,        sizeof(hparams.n_mels));
+        finp.read((char *) &hparams.f16,           sizeof(hparams.f16));
+
+        fprintf(stderr, "%s: n_vocab       = %d\n", __func__, hparams.n_vocab);
+        fprintf(stderr, "%s: n_audio_ctx   = %d\n", __func__, hparams.n_audio_ctx);
+        fprintf(stderr, "%s: n_audio_state = %d\n", __func__, hparams.n_audio_state);
+        fprintf(stderr, "%s: n_audio_head  = %d\n", __func__, hparams.n_audio_head);
+        fprintf(stderr, "%s: n_audio_layer = %d\n", __func__, hparams.n_audio_layer);
+        fprintf(stderr, "%s: n_text_ctx    = %d\n", __func__, hparams.n_text_ctx);
+        fprintf(stderr, "%s: n_text_state  = %d\n", __func__, hparams.n_text_state);
+        fprintf(stderr, "%s: n_text_head   = %d\n", __func__, hparams.n_text_head);
+        fprintf(stderr, "%s: n_text_layer  = %d\n", __func__, hparams.n_text_layer);
+        fprintf(stderr, "%s: n_mels        = %d\n", __func__, hparams.n_mels);
+        fprintf(stderr, "%s: f16           = %d\n", __func__, hparams.f16);
+
+        fout.write((char *) &hparams.n_vocab,       sizeof(hparams.n_vocab));
+        fout.write((char *) &hparams.n_audio_ctx,   sizeof(hparams.n_audio_ctx));
+        fout.write((char *) &hparams.n_audio_state, sizeof(hparams.n_audio_state));
+        fout.write((char *) &hparams.n_audio_head,  sizeof(hparams.n_audio_head));
+        fout.write((char *) &hparams.n_audio_layer, sizeof(hparams.n_audio_layer));
+        fout.write((char *) &hparams.n_text_ctx,    sizeof(hparams.n_text_ctx));
+        fout.write((char *) &hparams.n_text_state,  sizeof(hparams.n_text_state));
+        fout.write((char *) &hparams.n_text_head,   sizeof(hparams.n_text_head));
+        fout.write((char *) &hparams.n_text_layer,  sizeof(hparams.n_text_layer));
+        fout.write((char *) &hparams.n_mels,        sizeof(hparams.n_mels));
+        fout.write((char *) &itype,                 sizeof(hparams.f16));
+    }
+
+    // load mel filters
+    {
+        whisper_filters filters;
+
+        finp.read ((char *) &filters.n_mel, sizeof(filters.n_mel));
+        fout.write((char *) &filters.n_mel, sizeof(filters.n_mel));
+        finp.read ((char *) &filters.n_fft, sizeof(filters.n_fft));
+        fout.write((char *) &filters.n_fft, sizeof(filters.n_fft));
+
+        filters.data.resize(filters.n_mel * filters.n_fft);
+        finp.read ((char *) filters.data.data(), filters.data.size() * sizeof(float));
+        fout.write((char *) filters.data.data(), filters.data.size() * sizeof(float));
+    }
+
+    // load vocab
+    {
+        int32_t n_vocab = 0;
+        finp.read ((char *) &n_vocab, sizeof(n_vocab));
+        fout.write((char *) &n_vocab, sizeof(n_vocab));
+
+        //if (n_vocab != hparams.n_vocab) {
+        //    fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
+        //            __func__, fname_inp.c_str(), n_vocab, hparams.n_vocab);
+        //    return false;
+        //}
+
+        std::string word;
+        for (int i = 0; i < n_vocab; i++) {
+            uint32_t len;
+            finp.read ((char *) &len, sizeof(len));
+            fout.write((char *) &len, sizeof(len));
+
+            word.resize(len);
+            finp.read ((char *) word.data(), len);
+            fout.write((char *) word.data(), len);
+
+            vocab.token_to_id[word] = i;
+            vocab.id_to_token[i] = word;
+        }
+    }
+
+    // load weights
+    {
+        size_t total_size_org = 0;
+        size_t total_size_new = 0;
+
+        std::vector<float> work;
+
+        std::vector<uint8_t>     data_u8;
+        std::vector<ggml_fp16_t> data_f16;
+        std::vector<float>       data_f32;
+
+        std::vector<int64_t> hist_all(1 << 4, 0);
+
+        while (true) {
+            int32_t n_dims;
+            int32_t length;
+            int32_t ftype;
+
+            finp.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            finp.read(reinterpret_cast<char *>(&length), sizeof(length));
+            finp.read(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+
+            if (finp.eof()) {
+                break;
+            }
+
+            int32_t nelements = 1;
+            int32_t ne[3] = { 1, 1, 1 };
+            for (int i = 0; i < n_dims; ++i) {
+                finp.read (reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+                nelements *= ne[i];
+            }
+
+            std::string name(length, 0);
+            finp.read (&name[0], length);
+
+            {
+                static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
+                printf("%48s - [%5d, %5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ne[2], ftype_str[ftype]);
+            }
+
+            // regexes of tensor names to not be quantized
+            const std::vector<std::string> k_names = {
+                //"encoder.*",
+                "encoder.conv1.bias",
+                "encoder.conv2.bias",
+                "encoder.positional_embedding",
+                "decoder.positional_embedding",
+            };
+
+            bool quantize = true;
+            for (const auto & s : k_names) {
+                if (std::regex_match(name, std::regex(s))) {
+                    quantize = false;
+                    break;
+                }
+            }
+
+            // quantize only 2D and 3D tensors
+            quantize &= (n_dims == 2);
+
+            if (quantize) {
+                if (ftype != 0 && ftype != 1) {
+                    fprintf(stderr, "%s: unsupported ftype %d for integer quantization\n", __func__, ftype);
+                    return false;
+                }
+
+                if (ftype == 1) {
+                    data_f16.resize(nelements);
+                    finp.read(reinterpret_cast<char *>(data_f16.data()), nelements * sizeof(ggml_fp16_t));
+                    data_f32.resize(nelements);
+                    for (int i = 0; i < nelements; ++i) {
+                        data_f32[i] = ggml_fp16_to_fp32(data_f16[i]);
+                    }
+                } else {
+                    data_f32.resize(nelements);
+                    finp.read(reinterpret_cast<char *>(data_f32.data()), nelements * sizeof(float));
+                }
+
+                ftype = itype;
+            } else {
+                const int bpe = (ftype == 0) ? sizeof(float) : sizeof(uint16_t);
+
+                data_u8.resize(nelements*bpe);
+                finp.read(reinterpret_cast<char *>(data_u8.data()), nelements * bpe);
+            }
+
+            fout.write(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
+            fout.write(reinterpret_cast<char *>(&length), sizeof(length));
+            fout.write(reinterpret_cast<char *>(&ftype),  sizeof(ftype));
+            for (int i = 0; i < n_dims; ++i) {
+                fout.write(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
+            }
+            fout.write(&name[0], length);
+
+            if (quantize) {
+                printf("quantizing .. ");
+                work.resize(nelements); // for quantization
+
+                size_t cur_size = 0;
+                std::vector<int64_t> hist_cur(1 << 4, 0);
+
+                switch (type) {
+                    case GGML_TYPE_Q4_0:
+                        {
+                            cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], QK, hist_cur.data());
+                        } break;
+                    case GGML_TYPE_Q4_1:
+                        {
+                            cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], QK, hist_cur.data());
+                        } break;
+                    default:
+                        {
+                            fprintf(stderr, "%s: unsupported quantization type %d\n", __func__, type);
+                            return false;
+                        }
+                }
+
+                fout.write(reinterpret_cast<char *>(work.data()), cur_size);
+                total_size_new += cur_size;
+
+                printf("size = %8.3f MB -> %8.3f MB | hist: ", nelements * sizeof(float)/1024.0/1024.0, cur_size/1024.0/1024.0);
+                for (int i = 0; i < hist_cur.size(); ++i) {
+                    hist_all[i] += hist_cur[i];
+                }
+
+                for (int i = 0; i < hist_cur.size(); ++i) {
+                    printf("%5.3f ", hist_cur[i] / (float)nelements);
+                }
+                printf("\n");
+            } else {
+                printf("size = %8.3f MB\n", data_u8.size()/1024.0/1024.0);
+                fout.write(reinterpret_cast<char *>(data_u8.data()), data_u8.size());
+                total_size_new += data_u8.size();
+            }
+
+            total_size_org += nelements * sizeof(float);
+        }
+
+        printf("%s: model size  = %8.2f MB\n", __func__, total_size_org/1024.0/1024.0);
+        printf("%s: quant size  = %8.2f MB\n", __func__, total_size_new/1024.0/1024.0);
+
+        {
+            int64_t sum_all = 0;
+            for (int i = 0; i < hist_all.size(); ++i) {
+                sum_all += hist_all[i];
+            }
+
+            printf("%s: hist: ", __func__);
+            for (int i = 0; i < hist_all.size(); ++i) {
+                printf("%5.3f ", hist_all[i] / (float)sum_all);
+            }
+            printf("\n");
+        }
+    }
+
+    finp.close();
+    fout.close();
+
+    return true;
+}
+
+// usage:
+//  ./gpt-2-quantize models/gpt-2-117M/ggml-model.bin models/gpt-2-117M/ggml-model-quant.bin type
+//
+int main(int argc, char ** argv) {
+    if (argc != 4) {
+        fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
+        fprintf(stderr, "  type = 2 - q4_0\n");
+        fprintf(stderr, "  type = 3 - q4_1\n");
+        return 1;
+    }
+
+    // needed to initialize f16 tables
+    {
+        struct ggml_init_params params = { 0, NULL };
+        struct ggml_context * ctx = ggml_init(params);
+        ggml_free(ctx);
+    }
+
+    const std::string fname_inp = argv[1];
+    const std::string fname_out = argv[2];
+
+    const int itype = atoi(argv[3]);
+
+    const int64_t t_main_start_us = ggml_time_us();
+
+    int64_t t_quantize_us = 0;
+
+    // load the model
+    {
+        const int64_t t_start_us = ggml_time_us();
+
+        if (!whisper_model_quantize(fname_inp, fname_out, itype)) {
+            fprintf(stderr, "%s: failed to quantize model from '%s'\n", __func__, fname_inp.c_str());
+            return 1;
+        }
+
+        t_quantize_us = ggml_time_us() - t_start_us;
+    }
+
+    // report timing
+    {
+        const int64_t t_main_end_us = ggml_time_us();
+
+        printf("\n");
+        printf("%s: quantize time = %8.2f ms\n", __func__, t_quantize_us/1000.0f);
+        printf("%s:    total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
+    }
+
+    return 0;
+}

examples/whisper/whisper.cpp

@@ -252,12 +252,34 @@ static const std::map<e_model, size_t> MEM_REQ_SCRATCH3 = {
     { MODEL_LARGE,     9ull*MB },
 };
 
-static const std::map<e_model, size_t> MEM_REQ_MODEL = {
-    { MODEL_TINY,     74ull*MB },
-    { MODEL_BASE,    142ull*MB },
-    { MODEL_SMALL,   466ull*MB },
-    { MODEL_MEDIUM, 1464ull*MB },
-    { MODEL_LARGE,  2952ull*MB },
+static const std::map<ggml_type, std::map<e_model, size_t>> MEM_REQ_MODEL = {
+    { GGML_TYPE_F16,
+        {
+            { MODEL_TINY,     74ull*MB },
+            { MODEL_BASE,    142ull*MB },
+            { MODEL_SMALL,   466ull*MB },
+            { MODEL_MEDIUM, 1464ull*MB },
+            { MODEL_LARGE,  2952ull*MB },
+        },
+    },
+    { GGML_TYPE_Q4_0,
+        {
+            { MODEL_TINY,     26ull*MB },
+            { MODEL_BASE,     50ull*MB },
+            { MODEL_SMALL,   154ull*MB },
+            { MODEL_MEDIUM,  470ull*MB },
+            { MODEL_LARGE,   940ull*MB },
+        },
+    },
+    { GGML_TYPE_Q4_1,
+        {
+            { MODEL_TINY,     31ull*MB },
+            { MODEL_BASE,     57ull*MB },
+            { MODEL_SMALL,   181ull*MB },
+            { MODEL_MEDIUM,  559ull*MB },
+            { MODEL_LARGE,  1122ull*MB },
+        },
+    },
 };
 
 static const std::map<e_model, size_t> MEM_REQ_KV_SELF = {
@@ -681,7 +703,7 @@ static bool kv_cache_reinit(struct whisper_kv_cache & cache) {
     const ggml_type wtype = cache.k->type;
     WHISPER_ASSERT(wtype == cache.v->type);
 
-    WHISPER_ASSERT(cache.buf.size() >= 2*n_elements*ggml_type_size(wtype));
+    WHISPER_ASSERT(cache.buf.size() >= 2*n_elements*ggml_type_sizef(wtype));
 
     struct ggml_init_params params;
     params.mem_size   = cache.buf.size();
@@ -776,12 +798,25 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
             model.type = e_model::MODEL_LARGE;
         }
 
-        // for the big tensors, we have the option to store the data in 16-bit floats
+        // for the big tensors, we have the option to store the data in 16-bit floats or quantized
         // in order to save memory and also to speed up the computation
-        wctx.wtype = model.hparams.f16 ? GGML_TYPE_F16 : GGML_TYPE_F32;
+        wctx.wtype = GGML_TYPE_COUNT;
+        switch (model.hparams.f16) {
+            case 0: wctx.wtype = GGML_TYPE_F32;  break;
+            case 1: wctx.wtype = GGML_TYPE_F16;  break;
+            case 2: wctx.wtype = GGML_TYPE_Q4_0; break;
+            case 3: wctx.wtype = GGML_TYPE_Q4_1; break;
+            default:
+                    {
+                        fprintf(stderr, "%s: invalid model (bad f16 value %d)\n", __func__, model.hparams.f16);
+                        return false;
+                    }
+        }
 
         const size_t scale = model.hparams.f16 ? 1 : 2;
 
+        static const char * ftype_str[] = { "f32", "f16", "q4_0", "q4_1", };
+
         fprintf(stderr, "%s: n_vocab       = %d\n", __func__, hparams.n_vocab);
         fprintf(stderr, "%s: n_audio_ctx   = %d\n", __func__, hparams.n_audio_ctx);
         fprintf(stderr, "%s: n_audio_state = %d\n", __func__, hparams.n_audio_state);
@@ -792,7 +827,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
         fprintf(stderr, "%s: n_text_head   = %d\n", __func__, hparams.n_text_head);
         fprintf(stderr, "%s: n_text_layer  = %d\n", __func__, hparams.n_text_layer);
         fprintf(stderr, "%s: n_mels        = %d\n", __func__, hparams.n_mels);
-        fprintf(stderr, "%s: f16           = %d\n", __func__, hparams.f16);
+        fprintf(stderr, "%s: ftype         = %s\n", __func__, ftype_str[model.hparams.f16]);
         fprintf(stderr, "%s: type          = %d\n", __func__, model.type);
 
         // print memory requirements
@@ -803,7 +838,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
                      MEM_REQ_SCRATCH1.at (model.type) +
                      MEM_REQ_SCRATCH2.at (model.type) +
                      MEM_REQ_SCRATCH3.at (model.type) +
-                scale*MEM_REQ_MODEL.at   (model.type) +
+                scale*MEM_REQ_MODEL.at(wctx.wtype).at(model.type) +
                 scale*MEM_REQ_KV_CROSS.at(model.type) +
                 scale*std::max(MEM_REQ_ENCODE.at(model.type), MEM_REQ_DECODE.at(model.type));
 
@@ -819,9 +854,9 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
         // always have at least one decoder
 
         wctx.model.buf = new std::vector<uint8_t>();
-        wctx.model.buf->resize(scale*MEM_REQ_MODEL.at(model.type));
+        wctx.model.buf->resize(scale*MEM_REQ_MODEL.at(wctx.wtype).at(model.type));
 
-        if (!kv_cache_init(model.hparams, scale*MEM_REQ_KV_SELF.at(model.type), wctx.decoders[0].kv_self, wctx.wtype, model.hparams.n_text_ctx)) {
+        if (!kv_cache_init(model.hparams, scale*MEM_REQ_KV_SELF.at(model.type), wctx.decoders[0].kv_self, GGML_TYPE_F16, model.hparams.n_text_ctx)) {
             fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
             return false;
         }
@@ -831,7 +866,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
             fprintf(stderr, "%s: kv self size  = %7.2f MB\n", __func__, memory_size/1024.0/1024.0);
         }
 
-        if (!kv_cache_init(model.hparams, scale*MEM_REQ_KV_CROSS.at(model.type), wctx.kv_cross, wctx.wtype, model.hparams.n_audio_ctx)) {
+        if (!kv_cache_init(model.hparams, scale*MEM_REQ_KV_CROSS.at(model.type), wctx.kv_cross, GGML_TYPE_F16, model.hparams.n_audio_ctx)) {
             fprintf(stderr, "%s: kv_cache_init() failed for cross-attention cache\n", __func__);
             return false;
         }
@@ -963,92 +998,92 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
 
         // encoder
         {
-            ctx_size += n_audio_ctx*n_audio_state*ggml_type_size(GGML_TYPE_F32); // e_pe;
+            ctx_size += n_audio_ctx*n_audio_state*ggml_type_sizef(GGML_TYPE_F32); // e_pe;
 
-            ctx_size += 3*n_mels*n_audio_state*ggml_type_size(wtype);         // e_conv_1_w
-            ctx_size +=          n_audio_state*ggml_type_size(GGML_TYPE_F32); // e_conv_1_b
+            ctx_size += 3*n_mels*n_audio_state*ggml_type_sizef(GGML_TYPE_F16); // e_conv_1_w
+            ctx_size +=          n_audio_state*ggml_type_sizef(GGML_TYPE_F32); // e_conv_1_b
 
-            ctx_size += 3*n_audio_state*n_audio_state*ggml_type_size(wtype);         // e_conv_2_w
-            ctx_size +=                 n_audio_state*ggml_type_size(GGML_TYPE_F32); // e_conv_2_b
+            ctx_size += 3*n_audio_state*n_audio_state*ggml_type_sizef(GGML_TYPE_F16); // e_conv_2_w
+            ctx_size +=                 n_audio_state*ggml_type_sizef(GGML_TYPE_F32); // e_conv_2_b
 
-            ctx_size += n_audio_state*ggml_type_size(GGML_TYPE_F32); // e_ln_w;
-            ctx_size += n_audio_state*ggml_type_size(GGML_TYPE_F32); // e_ln_b;
+            ctx_size += n_audio_state*ggml_type_sizef(GGML_TYPE_F32); // e_ln_w;
+            ctx_size += n_audio_state*ggml_type_sizef(GGML_TYPE_F32); // e_ln_b;
         }
 
         // decoder
         {
-            ctx_size += n_text_ctx*n_text_state*ggml_type_size(GGML_TYPE_F32); // d_pe;
+            ctx_size += n_text_ctx*n_text_state*ggml_type_sizef(GGML_TYPE_F32); // d_pe;
 
-            ctx_size += n_vocab*n_text_state*ggml_type_size(wtype); // d_te;
+            ctx_size += n_vocab*n_text_state*ggml_type_sizef(wtype); // d_te;
 
-            ctx_size += n_text_state*ggml_type_size(GGML_TYPE_F32); // d_ln_w;
-            ctx_size += n_text_state*ggml_type_size(GGML_TYPE_F32); // d_ln_b;
+            ctx_size += n_text_state*ggml_type_sizef(GGML_TYPE_F32); // d_ln_w;
+            ctx_size += n_text_state*ggml_type_sizef(GGML_TYPE_F32); // d_ln_b;
         }
 
         // encoder layers
         {
-            ctx_size += n_audio_layer*(n_audio_state*ggml_type_size(GGML_TYPE_F32)); // mlp_ln_w
-            ctx_size += n_audio_layer*(n_audio_state*ggml_type_size(GGML_TYPE_F32)); // mlp_ln_b
+            ctx_size += n_audio_layer*(n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_ln_w
+            ctx_size += n_audio_layer*(n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_ln_b
 
-            ctx_size += n_audio_layer*(4*n_audio_state*n_audio_state*ggml_type_size(wtype));         // mlp_0_w
-            ctx_size += n_audio_layer*(              4*n_audio_state*ggml_type_size(GGML_TYPE_F32)); // mlp_0_b
+            ctx_size += n_audio_layer*(4*n_audio_state*n_audio_state*ggml_type_sizef(wtype));         // mlp_0_w
+            ctx_size += n_audio_layer*(              4*n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_0_b
 
-            ctx_size += n_audio_layer*(4*n_audio_state*n_audio_state*ggml_type_size(wtype));         // mlp_1_w
-            ctx_size += n_audio_layer*(                n_audio_state*ggml_type_size(GGML_TYPE_F32)); // mlp_1_b
+            ctx_size += n_audio_layer*(4*n_audio_state*n_audio_state*ggml_type_sizef(wtype));         // mlp_1_w
+            ctx_size += n_audio_layer*(                n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_1_b
 
-            ctx_size += n_audio_layer*(n_audio_state*ggml_type_size(GGML_TYPE_F32)); // attn_ln_0_w
-            ctx_size += n_audio_layer*(n_audio_state*ggml_type_size(GGML_TYPE_F32)); // attn_ln_0_b
+            ctx_size += n_audio_layer*(n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_ln_0_w
+            ctx_size += n_audio_layer*(n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_ln_0_b
 
-            ctx_size += n_audio_layer*(n_audio_state*n_audio_state*ggml_type_size(wtype));         // attn_q_w
-            ctx_size += n_audio_layer*(              n_audio_state*ggml_type_size(GGML_TYPE_F32)); // attn_q_b
+            ctx_size += n_audio_layer*(n_audio_state*n_audio_state*ggml_type_sizef(wtype));         // attn_q_w
+            ctx_size += n_audio_layer*(              n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_q_b
 
-            ctx_size += n_audio_layer*(n_audio_state*n_audio_state*ggml_type_size(wtype)); // attn_k_w
+            ctx_size += n_audio_layer*(n_audio_state*n_audio_state*ggml_type_sizef(wtype)); // attn_k_w
 
-            ctx_size += n_audio_layer*(n_audio_state*n_audio_state*ggml_type_size(wtype));         // attn_v_w
-            ctx_size += n_audio_layer*(              n_audio_state*ggml_type_size(GGML_TYPE_F32)); // attn_v_b
+            ctx_size += n_audio_layer*(n_audio_state*n_audio_state*ggml_type_sizef(wtype));         // attn_v_w
+            ctx_size += n_audio_layer*(              n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_v_b
 
-            ctx_size += n_audio_layer*(n_audio_state*n_audio_state*ggml_type_size(wtype));         // attn_ln_1_w
-            ctx_size += n_audio_layer*(              n_audio_state*ggml_type_size(GGML_TYPE_F32)); // attn_ln_1_b
+            ctx_size += n_audio_layer*(n_audio_state*n_audio_state*ggml_type_sizef(wtype));         // attn_ln_1_w
+            ctx_size += n_audio_layer*(              n_audio_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_ln_1_b
         }
 
         // decoder layers
         {
-            ctx_size += n_text_layer*(n_text_state*ggml_type_size(GGML_TYPE_F32)); // mlp_ln_w
-            ctx_size += n_text_layer*(n_text_state*ggml_type_size(GGML_TYPE_F32)); // mlp_ln_b
+            ctx_size += n_text_layer*(n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_ln_w
+            ctx_size += n_text_layer*(n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_ln_b
 
-            ctx_size += n_text_layer*(4*n_text_state*n_text_state*ggml_type_size(wtype));         // mlp_0_w
-            ctx_size += n_text_layer*(             4*n_text_state*ggml_type_size(GGML_TYPE_F32)); // mlp_0_b
+            ctx_size += n_text_layer*(4*n_text_state*n_text_state*ggml_type_sizef(wtype));         // mlp_0_w
+            ctx_size += n_text_layer*(             4*n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_0_b
 
-            ctx_size += n_text_layer*(4*n_text_state*n_text_state*ggml_type_size(wtype));         // mlp_1_w
-            ctx_size += n_text_layer*(               n_text_state*ggml_type_size(GGML_TYPE_F32)); // mlp_1_b
+            ctx_size += n_text_layer*(4*n_text_state*n_text_state*ggml_type_sizef(wtype));         // mlp_1_w
+            ctx_size += n_text_layer*(               n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // mlp_1_b
 
-            ctx_size += n_text_layer*(n_text_state*ggml_type_size(GGML_TYPE_F32)); // attn_ln_0_w
-            ctx_size += n_text_layer*(n_text_state*ggml_type_size(GGML_TYPE_F32)); // attn_ln_0_b
+            ctx_size += n_text_layer*(n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_ln_0_w
+            ctx_size += n_text_layer*(n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_ln_0_b
 
-            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_size(wtype));         // attn_q_w
-            ctx_size += n_text_layer*(             n_text_state*ggml_type_size(GGML_TYPE_F32)); // attn_q_b
+            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype));         // attn_q_w
+            ctx_size += n_text_layer*(             n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_q_b
 
-            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_size(wtype)); // attn_k_w
+            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype)); // attn_k_w
 
-            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_size(wtype));         // attn_v_w
-            ctx_size += n_text_layer*(             n_text_state*ggml_type_size(GGML_TYPE_F32)); // attn_v_b
+            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype));         // attn_v_w
+            ctx_size += n_text_layer*(             n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_v_b
 
-            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_size(wtype));         // attn_ln_1_w
-            ctx_size += n_text_layer*(             n_text_state*ggml_type_size(GGML_TYPE_F32)); // attn_ln_1_b
+            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype));         // attn_ln_1_w
+            ctx_size += n_text_layer*(             n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // attn_ln_1_b
                                                                                                 //
-            ctx_size += n_text_layer*(n_text_state*ggml_type_size(GGML_TYPE_F32)); // cross_attn_ln_0_w
-            ctx_size += n_text_layer*(n_text_state*ggml_type_size(GGML_TYPE_F32)); // cross_attn_ln_0_b
+            ctx_size += n_text_layer*(n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // cross_attn_ln_0_w
+            ctx_size += n_text_layer*(n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // cross_attn_ln_0_b
 
-            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_size(wtype));         // cross_attn_q_w
-            ctx_size += n_text_layer*(             n_text_state*ggml_type_size(GGML_TYPE_F32)); // cross_attn_q_b
+            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype));         // cross_attn_q_w
+            ctx_size += n_text_layer*(             n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // cross_attn_q_b
 
-            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_size(wtype)); // cross_attn_k_w
+            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype)); // cross_attn_k_w
 
-            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_size(wtype));         // cross_attn_v_w
-            ctx_size += n_text_layer*(             n_text_state*ggml_type_size(GGML_TYPE_F32)); // cross_attn_v_b
+            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype));         // cross_attn_v_w
+            ctx_size += n_text_layer*(             n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // cross_attn_v_b
 
-            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_size(wtype));         // cross_attn_ln_1_w
-            ctx_size += n_text_layer*(             n_text_state*ggml_type_size(GGML_TYPE_F32)); // cross_attn_ln_1_b
+            ctx_size += n_text_layer*(n_text_state*n_text_state*ggml_type_sizef(wtype));         // cross_attn_ln_1_w
+            ctx_size += n_text_layer*(             n_text_state*ggml_type_sizef(GGML_TYPE_F32)); // cross_attn_ln_1_b
         }
 
         ctx_size += (15 + 15*n_audio_layer + 24*n_text_layer)*256; // object overhead
@@ -1094,10 +1129,10 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
         {
             model.e_pe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_audio_state, n_audio_ctx);
 
-            model.e_conv_1_w = ggml_new_tensor_3d(ctx, wtype,         3, n_mels, n_audio_state);
+            model.e_conv_1_w = ggml_new_tensor_3d(ctx, GGML_TYPE_F16, 3, n_mels, n_audio_state);
             model.e_conv_1_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, n_audio_state);
 
-            model.e_conv_2_w = ggml_new_tensor_3d(ctx, wtype,         3, n_audio_state, n_audio_state);
+            model.e_conv_2_w = ggml_new_tensor_3d(ctx, GGML_TYPE_F16, 3, n_audio_state, n_audio_state);
             model.e_conv_2_b = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, 1, n_audio_state);
 
             model.e_ln_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
@@ -1313,9 +1348,21 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
                 return false;
             }
 
-            const size_t bpe = (ftype == 0) ? sizeof(float) : sizeof(ggml_fp16_t);
+            size_t bpe = 0;
+
+            switch (ftype) {
+                case 0: bpe = ggml_type_size(GGML_TYPE_F32);  break;
+                case 1: bpe = ggml_type_size(GGML_TYPE_F16);  break;
+                case 2: bpe = ggml_type_size(GGML_TYPE_Q4_0); assert(ne[0] % 64 == 0); break;
+                case 3: bpe = ggml_type_size(GGML_TYPE_Q4_1); assert(ne[0] % 64 == 0); break;
+                default:
+                        {
+                            fprintf(stderr, "%s: unknown ftype %d in model file\n", __func__, ftype);
+                            return false;
+                        }
+            };
 
-            if (nelements*bpe != ggml_nbytes(tensor)) {
+            if ((nelements*bpe)/ggml_blck_size(tensor->type) != ggml_nbytes(tensor)) {
                 fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
                         __func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
                 return false;
@@ -1513,14 +1560,14 @@ static bool whisper_encode(
                 ggml_permute(ctx0,
                         ggml_cpy(ctx0,
                             Qcur,
-                            ggml_new_tensor_3d(ctx0, wctx.wtype, n_state/n_head, n_head, n_ctx)),
+                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F16, n_state/n_head, n_head, n_ctx)),
                         0, 2, 1, 3);
 
             struct ggml_tensor * K =
                 ggml_permute(ctx0,
                         ggml_cpy(ctx0,
                             Kcur,
-                            ggml_new_tensor_3d(ctx0, wctx.wtype, n_state/n_head, n_head, n_ctx)),
+                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F16, n_state/n_head, n_head, n_ctx)),
                         0, 2, 1, 3);
 
             struct ggml_tensor * V =
@@ -1530,7 +1577,7 @@ static bool whisper_encode(
                                 Vcur,
                                 n_state/n_head, n_head, n_ctx),
                             1, 2, 0, 3),
-                        ggml_new_tensor_3d(ctx0, wctx.wtype, n_ctx, n_state/n_head, n_head)
+                        ggml_new_tensor_3d(ctx0, GGML_TYPE_F16, n_ctx, n_state/n_head, n_head)
                         );
 
             struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, false);
@@ -1546,7 +1593,7 @@ static bool whisper_encode(
                 ggml_permute(ctx0,
                         ggml_cpy(ctx0,
                             Kcur,
-                            ggml_new_tensor_3d(ctx0, wctx.wtype, n_state/n_head, n_head, n_ctx)),
+                            ggml_new_tensor_3d(ctx0, GGML_TYPE_F16, n_state/n_head, n_head, n_ctx)),
                         0, 2, 1, 3);
 
             // K * Q
@@ -1564,7 +1611,7 @@ static bool whisper_encode(
             //    ggml_permute(ctx0,
             //            ggml_cpy(ctx0,
             //                Vcur,
-            //                ggml_new_tensor_3d(ctx0, wctx.wtype, n_state/n_head, n_head, n_ctx)),
+            //                ggml_new_tensor_3d(ctx0, GGML_TYPE_F16, n_state/n_head, n_head, n_ctx)),
             //            1, 2, 0, 3);
 
             //struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
@@ -1576,7 +1623,7 @@ static bool whisper_encode(
                                 Vcur,
                                 n_state/n_head, n_head, n_ctx),
                             0, 2, 1, 3),
-                        ggml_new_tensor_3d(ctx0, wctx.wtype, n_state/n_head, n_ctx, n_head)
+                        ggml_new_tensor_3d(ctx0, GGML_TYPE_F16, n_state/n_head, n_ctx, n_head)
                         );
 
             struct ggml_tensor * KQV = ggml_mul_mat(ctx0, ggml_transpose(ctx0, V), KQ_soft_max);
@@ -1634,7 +1681,7 @@ static bool whisper_encode(
             wctx.use_buf(ctx0, 0);
 
             cur = ggml_flash_ff(ctx0,
-                    ggml_cpy(ctx0, cur, ggml_new_tensor_2d(ctx0, wctx.wtype, n_state, n_ctx)),
+                    ggml_cpy(ctx0, cur, ggml_new_tensor_2d(ctx0, GGML_TYPE_F16, n_state, n_ctx)),
                     layer.mlp_0_w, layer.mlp_0_b, layer.mlp_1_w, layer.mlp_1_b);
 #else
             wctx.use_buf(ctx0, 0);

include/ggml/ggml.h

@@ -198,6 +198,8 @@ struct ggml_object;
 struct ggml_context;
 
 enum ggml_type {
+    GGML_TYPE_Q4_0,
+    GGML_TYPE_Q4_1,
     GGML_TYPE_I8,
     GGML_TYPE_I16,
     GGML_TYPE_I32,
@@ -226,6 +228,7 @@ enum ggml_op {
     GGML_OP_STEP,
     GGML_OP_RELU,
     GGML_OP_GELU,
+    GGML_OP_SILU,
     GGML_OP_NORM, // normalize
 
     GGML_OP_MUL_MAT,
@@ -326,7 +329,10 @@ void ggml_print_objects(const struct ggml_context * ctx);
 int    ggml_nelements(const struct ggml_tensor * tensor);
 size_t ggml_nbytes   (const struct ggml_tensor * tensor);
 
-size_t ggml_type_size   (enum ggml_type type);
+int    ggml_blck_size (enum ggml_type type);
+size_t ggml_type_size (enum ggml_type type); // size in bytes for all elements in a block
+float  ggml_type_sizef(enum ggml_type type); // ggml_type_size()/ggml_blck_size() as float
+
 size_t ggml_element_size(const struct ggml_tensor * tensor);
 
 struct ggml_context * ggml_init(struct ggml_init_params params);
@@ -466,6 +472,10 @@ struct ggml_tensor * ggml_gelu(
         struct ggml_context * ctx,
         struct ggml_tensor  * a);
 
+struct ggml_tensor * ggml_silu(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
 // normalize along rows
 // TODO: eps is hardcoded to 1e-5 for now
 struct ggml_tensor * ggml_norm(