Initial release

2022-09-25 21:23:15 +03:00 · 2022-09-25 21:23:15 +03:00 · b0a11594ae
commit b0a11594ae
10 changed files with 16208 additions and 0 deletions
--- a/.gitignore
+++ b/.gitignore
@ -0,0 +1,3 @@
+sync.sh
+main
+*.o
--- a/109
+++ b/109
@ -0,0 +1,109 @@
+main: ggml.o main.o
+	g++ -o main ggml.o main.o
+
+ggml.o: ggml.c ggml.h
+	gcc -O3 -mavx -mavx2 -mfma -mf16c -c ggml.c
+
+main.o: main.cpp ggml.h
+	g++ -O3 -std=c++11 -c main.cpp
+
+# clean up the directory
+clean:
+	rm -f *.o main
+
+# run the program
+run: main
+	./main
+
+# download the following audio samples into folder "./samples":
+.PHONY: samples
+samples:
+	@echo "Downloading samples..."
+	mkdir -p samples
+	@wget --quiet --show-progress -O samples/gb0.ogg https://upload.wikimedia.org/wikipedia/commons/2/22/George_W._Bush%27s_weekly_radio_address_%28November_1%2C_2008%29.oga
+	@wget --quiet --show-progress -O samples/gb1.ogg https://upload.wikimedia.org/wikipedia/commons/1/1f/George_W_Bush_Columbia_FINAL.ogg
+	@wget --quiet --show-progress -O samples/hp0.ogg https://upload.wikimedia.org/wikipedia/en/d/d4/En.henryfphillips.ogg
+	@echo "Converting to 16-bit WAV ..."
+	@ffmpeg -loglevel -0 -y -i samples/gb0.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/gb0.wav
+	@ffmpeg -loglevel -0 -y -i samples/gb1.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/gb1.wav
+	@ffmpeg -loglevel -0 -y -i samples/hp0.ogg -ar 16000 -ac 1 -c:a pcm_s16le samples/hp0.wav
+
+.PHONY: tiny.en
+tiny.en: main
+	@echo "Downloading tiny.en (75 MB just once)"
+	mkdir -p models
+	@if [ ! -f models/ggml-tiny.en.bin ]; then \
+		wget --quiet --show-progress -O models/ggml-tiny.en.bin https://ggml.ggerganov.com/ggml-model-whisper-tiny.en.bin ; \
+	fi
+	@echo "==============================================="
+	@echo "Running tiny.en on all samples in ./samples ..."
+	@echo "==============================================="
+	@echo ""
+	@for f in samples/*.wav; do \
+		echo "----------------------------------------------" ; \
+		echo "[+] Running base.en on $$f ... (run 'ffplay $$f' to listen)" ; \
+	    echo "----------------------------------------------" ; \
+		echo "" ; \
+		./main -m models/ggml-tiny.en.bin -f $$f ; \
+		echo "" ; \
+	done
+
+.PHONY: base.en
+base.en: main
+	@echo "Downloading base.en (142 MB just once)"
+	mkdir -p models
+	@if [ ! -f models/ggml-base.en.bin ]; then \
+		wget --quiet --show-progress -O models/ggml-base.en.bin https://ggml.ggerganov.com/ggml-model-whisper-base.en.bin ; \
+	fi
+	@echo "==============================================="
+	@echo "Running base.en on all samples in ./samples ..."
+	@echo "==============================================="
+	@echo ""
+	@for f in samples/*.wav; do \
+		echo "----------------------------------------------" ; \
+		echo "[+] Running base.en on $$f ... (run 'ffplay $$f' to listen)" ; \
+	    echo "----------------------------------------------" ; \
+		echo "" ; \
+		./main -m models/ggml-base.en.bin -f $$f ; \
+		echo "" ; \
+	done
+
+.PHONY: small.en
+small.en: main
+	@echo "Downloading small.en (466 MB just once)"
+	mkdir -p models
+	@if [ ! -f models/ggml-small.en.bin ]; then \
+		wget --quiet --show-progress -O models/ggml-small.en.bin https://ggml.ggerganov.com/ggml-model-whisper-small.en.bin ; \
+	fi
+	@echo "==============================================="
+	@echo "Running small.en on all samples in ./samples ..."
+	@echo "==============================================="
+	@echo ""
+	@for f in samples/*.wav; do \
+		echo "----------------------------------------------" ; \
+		echo "[+] Running base.en on $$f ... (run 'ffplay $$f' to listen)" ; \
+	    echo "----------------------------------------------" ; \
+		echo "" ; \
+		./main -m models/ggml-small.en.bin -f $$f ; \
+		echo "" ; \
+	done
+
+.PHONY: medium.en
+medium.en: main
+	@echo "Downloading medium.en (1.5 GB just once)"
+	mkdir -p models
+	@if [ ! -f models/ggml-medium.en.bin ]; then \
+		wget --quiet --show-progress -O models/ggml-medium.en.bin https://ggml.ggerganov.com/ggml-model-whisper-medium.en.bin ; \
+	fi
+	@echo "==============================================="
+	@echo "Running medium.en on all samples in ./samples ..."
+	@echo "==============================================="
+	@echo ""
+	@for f in samples/*.wav; do \
+		echo "----------------------------------------------" ; \
+		echo "[+] Running base.en on $$f ... (run 'ffplay $$f' to listen)" ; \
+	    echo "----------------------------------------------" ; \
+		echo "" ; \
+		./main -m models/ggml-medium.en.bin -f $$f ; \
+		echo "" ; \
+	done
--- a/convert-pt-to-ggml.py
+++ b/convert-pt-to-ggml.py
@ -0,0 +1,328 @@
+# Convert Whisper transformer model from PyTorch to ggml format
+#
+# Usage: python convert-pt-to-ggml.py ~/.cache/whisper/medium.pt ~/path/to/repo/whisper/ ./models/whisper-medium
+#
+# You need to clone the original repo in ~/path/to/repo/whisper/
+#
+#  git clone https://github.com/openai/whisper ~/path/to/repo/whisper/
+#
+# It is used to various assets needed by the algorithm:
+#
+#  - tokenizer
+#  - mel filters
+#
+# Also, you need to have the original models in ~/.cache/whisper/
+# See the original repo for more details.
+#
+# This script loads the specified model and whisper assets and saves them in ggml format.
+# The output is a single binary file containing the following information:
+#
+#  - hparams
+#  - mel filters
+#  - tokenizer vocab
+#  - model variables
+#
+# 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])
+#
+
+import io
+import os
+import sys
+import struct
+import json
+import code
+import torch
+import numpy as np
+
+from transformers import GPTJForCausalLM
+from transformers import GPT2TokenizerFast
+
+# ref: https://github.com/openai/whisper/blob/8cf36f3508c9acd341a45eb2364239a3d81458b9/whisper/tokenizer.py#L10-L110
+LANGUAGES = {
+    "en": "english",
+    "zh": "chinese",
+    "de": "german",
+    "es": "spanish",
+    "ru": "russian",
+    "ko": "korean",
+    "fr": "french",
+    "ja": "japanese",
+    "pt": "portuguese",
+    "tr": "turkish",
+    "pl": "polish",
+    "ca": "catalan",
+    "nl": "dutch",
+    "ar": "arabic",
+    "sv": "swedish",
+    "it": "italian",
+    "id": "indonesian",
+    "hi": "hindi",
+    "fi": "finnish",
+    "vi": "vietnamese",
+    "iw": "hebrew",
+    "uk": "ukrainian",
+    "el": "greek",
+    "ms": "malay",
+    "cs": "czech",
+    "ro": "romanian",
+    "da": "danish",
+    "hu": "hungarian",
+    "ta": "tamil",
+    "no": "norwegian",
+    "th": "thai",
+    "ur": "urdu",
+    "hr": "croatian",
+    "bg": "bulgarian",
+    "lt": "lithuanian",
+    "la": "latin",
+    "mi": "maori",
+    "ml": "malayalam",
+    "cy": "welsh",
+    "sk": "slovak",
+    "te": "telugu",
+    "fa": "persian",
+    "lv": "latvian",
+    "bn": "bengali",
+    "sr": "serbian",
+    "az": "azerbaijani",
+    "sl": "slovenian",
+    "kn": "kannada",
+    "et": "estonian",
+    "mk": "macedonian",
+    "br": "breton",
+    "eu": "basque",
+    "is": "icelandic",
+    "hy": "armenian",
+    "ne": "nepali",
+    "mn": "mongolian",
+    "bs": "bosnian",
+    "kk": "kazakh",
+    "sq": "albanian",
+    "sw": "swahili",
+    "gl": "galician",
+    "mr": "marathi",
+    "pa": "punjabi",
+    "si": "sinhala",
+    "km": "khmer",
+    "sn": "shona",
+    "yo": "yoruba",
+    "so": "somali",
+    "af": "afrikaans",
+    "oc": "occitan",
+    "ka": "georgian",
+    "be": "belarusian",
+    "tg": "tajik",
+    "sd": "sindhi",
+    "gu": "gujarati",
+    "am": "amharic",
+    "yi": "yiddish",
+    "lo": "lao",
+    "uz": "uzbek",
+    "fo": "faroese",
+    "ht": "haitian creole",
+    "ps": "pashto",
+    "tk": "turkmen",
+    "nn": "nynorsk",
+    "mt": "maltese",
+    "sa": "sanskrit",
+    "lb": "luxembourgish",
+    "my": "myanmar",
+    "bo": "tibetan",
+    "tl": "tagalog",
+    "mg": "malagasy",
+    "as": "assamese",
+    "tt": "tatar",
+    "haw": "hawaiian",
+    "ln": "lingala",
+    "ha": "hausa",
+    "ba": "bashkir",
+    "jw": "javanese",
+    "su": "sundanese",
+}
+
+# ref: https://github.com/openai/whisper/blob/8cf36f3508c9acd341a45eb2364239a3d81458b9/whisper/tokenizer.py#L273-L292
+def build_tokenizer(path_to_whisper_repo: str, name: str = "gpt2"):
+    os.environ["TOKENIZERS_PARALLELISM"] = "false"
+    path = os.path.join(path_to_whisper_repo, "whisper/assets", name)
+    tokenizer = GPT2TokenizerFast.from_pretrained(path)
+
+    specials = [
+        "<|startoftranscript|>",
+        *[f"<|{lang}|>" for lang in LANGUAGES.keys()],
+        "<|translate|>",
+        "<|transcribe|>",
+        "<|startoflm|>",
+        "<|startofprev|>",
+        "<|nocaptions|>",
+        "<|notimestamps|>",
+    ]
+
+    tokenizer.add_special_tokens(dict(additional_special_tokens=specials))
+    return tokenizer
+
+# ref: https://github.com/openai/gpt-2/blob/master/src/encoder.py
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8+n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+if len(sys.argv) < 4:
+    print("Usage: convert-pt-to-ggml.py model.pt path-to-whisper-repo dir-output [use-f32]\n")
+    sys.exit(1)
+
+fname_inp   = sys.argv[1]
+dir_whisper = sys.argv[2]
+dir_out     = sys.argv[3]
+
+# try to load PyTorch binary data
+try:
+    model_bytes = open(fname_inp, "rb").read()
+    with io.BytesIO(model_bytes) as fp:
+        checkpoint = torch.load(fp, map_location="cpu")
+except:
+    print("Error: failed to load PyTorch model file: %s" % fname_inp)
+    sys.exit(1)
+
+hparams = checkpoint["dims"]
+print("hparams:", hparams)
+
+list_vars = checkpoint["model_state_dict"]
+
+#print(list_vars['encoder.positional_embedding'])
+#print(list_vars['encoder.conv1.weight'])
+#print(list_vars['encoder.conv1.weight'].shape)
+
+# load mel filters
+n_mels = hparams["n_mels"]
+with np.load(os.path.join(dir_whisper, "whisper/assets", "mel_filters.npz")) as f:
+    filters = torch.from_numpy(f[f"mel_{n_mels}"])
+    #print (filters)
+
+#code.interact(local=locals())
+
+multilingual = hparams["n_vocab"] == 51865
+tokenizer = build_tokenizer(dir_whisper, multilingual and "multilingual" or "gpt2")
+
+#print(tokenizer)
+#print(tokenizer.name_or_path)
+#print(len(tokenizer.additional_special_tokens))
+dir_tokenizer = tokenizer.name_or_path
+
+# output in the same directory as the model
+fname_out = dir_out + "/ggml-model.bin"
+
+with open(dir_tokenizer + "/vocab.json", "r") as f:
+    tokens = json.load(f)
+
+# use 16-bit or 32-bit floats
+use_f16 = True
+if len(sys.argv) > 4:
+    use_f16 = False
+    fname_out = dir_out + "/ggml-model-f32.bin"
+
+fout = open(fname_out, "wb")
+
+fout.write(struct.pack("i", 0x67676d6c)) # magic: ggml in hex
+fout.write(struct.pack("i", hparams["n_vocab"]))
+fout.write(struct.pack("i", hparams["n_audio_ctx"]))
+fout.write(struct.pack("i", hparams["n_audio_state"]))
+fout.write(struct.pack("i", hparams["n_audio_head"]))
+fout.write(struct.pack("i", hparams["n_audio_layer"]))
+fout.write(struct.pack("i", hparams["n_text_ctx"]))
+fout.write(struct.pack("i", hparams["n_text_state"]))
+fout.write(struct.pack("i", hparams["n_text_head"]))
+fout.write(struct.pack("i", hparams["n_text_layer"]))
+fout.write(struct.pack("i", hparams["n_mels"]))
+fout.write(struct.pack("i", use_f16))
+
+# write mel filters
+fout.write(struct.pack("i", filters.shape[0]))
+fout.write(struct.pack("i", filters.shape[1]))
+for i in range(filters.shape[0]):
+    for j in range(filters.shape[1]):
+        fout.write(struct.pack("f", filters[i][j]))
+
+byte_encoder = bytes_to_unicode()
+byte_decoder = {v:k for k, v in byte_encoder.items()}
+
+fout.write(struct.pack("i", len(tokens)))
+
+for key in tokens:
+    text = bytearray([byte_decoder[c] for c in key]).decode('utf-8', errors='replace').encode('utf-8')
+    fout.write(struct.pack("i", len(text)))
+    fout.write(text)
+
+for name in list_vars.keys():
+    data = list_vars[name].squeeze().numpy()
+    print("Processing variable: " + name + " with shape: ", data.shape)
+
+    # reshape conv bias from [n] to [n, 1]
+    if name == "encoder.conv1.bias" or \
+       name == "encoder.conv2.bias":
+        data = data.reshape(data.shape[0], 1)
+        print("  Reshaped variable: " + name + " to shape: ", data.shape)
+
+    n_dims = len(data.shape);
+
+    # looks like the whisper models are in f16 by default
+    # so we need to convert the small tensors to f32 until we fully support f16 in ggml
+    # ftype == 0 -> float32, ftype == 1 -> float16
+    ftype = 1;
+    if use_f16:
+        if n_dims < 2 or \
+                name == "encoder.conv1.bias"   or \
+                name == "encoder.conv2.bias"   or \
+                name == "encoder.positional_embedding" or \
+                name == "decoder.positional_embedding":
+            ftype = 0
+            data = data.astype(np.float32)
+            print("  Converting to float32")
+            data = data.astype(np.float32)
+            ftype = 0
+    else:
+        data = data.astype(np.float32)
+        ftype = 0
+
+    #if name.startswith("encoder"):
+    #    if name.endswith("mlp.0.weight") or \
+    #       name.endswith("mlp.2.weight"):
+    #        print("  Transposing")
+    #        data = data.transpose()
+
+    # header
+    str = name.encode('utf-8')
+    fout.write(struct.pack("iii", n_dims, len(str), ftype))
+    for i in range(n_dims):
+        fout.write(struct.pack("i", data.shape[n_dims - 1 - i]))
+    fout.write(str);
+
+    # data
+    data.tofile(fout)
+
+fout.close()
+
+print("Done. Output file: " + fname_out)
+print("")
--- a/dr_wav.h
+++ b/dr_wav.h
--- a/ggml.c
+++ b/ggml.c
--- a/ggml.h
+++ b/ggml.h
@ -0,0 +1,527 @@
+#pragma once
+
+#ifdef  __cplusplus
+extern "C" {
+#endif
+
+#include <stdint.h>
+#include <stddef.h>
+#include <stdbool.h>
+
+#define GGML_MAX_DIMS     4
+#define GGML_MAX_NODES    4096
+#define GGML_MAX_PARAMS   16
+#define GGML_MAX_CONTEXTS 16
+
+#ifdef __ARM_NEON
+// we use the built-in 16-bit float type
+typedef __fp16 ggml_fp16_t;
+#else
+typedef uint16_t ggml_fp16_t;
+#endif
+
+float ggml_fp16_to_fp32(ggml_fp16_t x);
+ggml_fp16_t ggml_fp32_to_fp16(float x);
+
+struct ggml_object;
+struct ggml_context;
+
+enum ggml_type {
+    GGML_TYPE_I8,
+    GGML_TYPE_I16,
+    GGML_TYPE_I32,
+    GGML_TYPE_F16,
+    GGML_TYPE_F32,
+    GGML_TYPE_COUNT,
+};
+
+enum ggml_op {
+    GGML_OP_NONE = 0,
+
+    GGML_OP_DUP,
+    GGML_OP_ADD,
+    GGML_OP_SUB,
+    GGML_OP_MUL,
+    GGML_OP_DIV,
+    GGML_OP_SQR,
+    GGML_OP_SQRT,
+    GGML_OP_SUM,
+    GGML_OP_MEAN,
+    GGML_OP_REPEAT,
+    GGML_OP_ABS,
+    GGML_OP_SGN,
+    GGML_OP_NEG,
+    GGML_OP_STEP,
+    GGML_OP_RELU,
+    GGML_OP_GELU,
+    GGML_OP_NORM, // normalize
+
+    GGML_OP_MUL_MAT,
+
+    GGML_OP_SCALE,
+    GGML_OP_CPY,
+    GGML_OP_RESHAPE,
+    GGML_OP_VIEW,
+    GGML_OP_PERMUTE,
+    GGML_OP_TRANSPOSE,
+    GGML_OP_GET_ROWS,
+    GGML_OP_DIAG_MASK_INF,
+    GGML_OP_SOFT_MAX,
+    GGML_OP_ROPE,
+    GGML_OP_CONV_1D_1S,
+    GGML_OP_CONV_1D_2S,
+
+    GGML_OP_COUNT,
+};
+
+// n-dimensional tensor
+struct ggml_tensor {
+    enum ggml_type type;
+
+    int    n_dims;
+    int    ne[GGML_MAX_DIMS]; // number of elements
+    size_t nb[GGML_MAX_DIMS]; // stride in bytes:
+                              // nb[0] = sizeof(type)
+                              // nb[1] = nb[0]   * ne[0] + padding
+                              // nb[i] = nb[i-1] * ne[i-1]
+
+    // compute data
+    enum ggml_op op;
+
+    bool is_param;
+
+    struct ggml_tensor * grad;
+    struct ggml_tensor * src0;
+    struct ggml_tensor * src1;
+
+    // thread scheduling
+    int n_tasks;
+
+    // performance
+    int     perf_runs;
+    int64_t perf_cycles;
+    int64_t perf_time_us;
+
+    void * data;
+    char pad[8];
+};
+
+// computation graph
+struct ggml_cgraph {
+    int n_nodes;
+    int n_leafs;
+    int n_threads;
+
+    size_t work_size;
+    struct ggml_tensor * work;
+
+    struct ggml_tensor * nodes[GGML_MAX_NODES];
+    struct ggml_tensor * grads[GGML_MAX_NODES];
+    struct ggml_tensor * leafs[GGML_MAX_NODES];
+
+    // performance
+    int     perf_runs;
+    int64_t perf_cycles;
+    int64_t perf_time_us;
+};
+
+struct ggml_init_params {
+    // memory pool
+    size_t mem_size;   // bytes
+    void * mem_buffer; // if NULL, memory will be allocated internally
+};
+
+int64_t ggml_time_ms(void);
+int64_t ggml_time_us(void);
+int64_t ggml_cycles(void);
+int64_t ggml_cycles_per_ms(void);
+
+void ggml_print_object (const struct ggml_object * obj);
+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);
+size_t ggml_element_size(const struct ggml_tensor * tensor);
+
+struct ggml_context * ggml_init(struct ggml_init_params params);
+void ggml_free(struct ggml_context * ctx);
+
+size_t ggml_used_mem(const struct ggml_context * ctx);
+
+struct ggml_tensor * ggml_new_tensor(
+        struct ggml_context * ctx,
+        enum   ggml_type type,
+        int    n_dims,
+        const int *ne);
+
+struct ggml_tensor * ggml_new_tensor_1d(
+        struct ggml_context * ctx,
+        enum   ggml_type type,
+        int    ne0);
+
+struct ggml_tensor * ggml_new_tensor_2d(
+        struct ggml_context * ctx,
+        enum   ggml_type type,
+        int    ne0,
+        int    ne1);
+
+struct ggml_tensor * ggml_new_tensor_3d(
+        struct ggml_context * ctx,
+        enum   ggml_type type,
+        int    ne0,
+        int    ne1,
+        int    ne2);
+
+struct ggml_tensor * ggml_new_tensor_4d(
+        struct ggml_context * ctx,
+        enum   ggml_type type,
+        int    ne0,
+        int    ne1,
+        int    ne2,
+        int    ne3);
+
+struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
+
+struct ggml_tensor * ggml_dup_tensor (struct ggml_context * ctx, const struct ggml_tensor * src);
+struct ggml_tensor * ggml_view_tensor(struct ggml_context * ctx, const struct ggml_tensor * src);
+
+struct ggml_tensor * ggml_set_zero(struct ggml_tensor * tensor);
+struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
+
+float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
+void  ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
+
+ void * ggml_get_data    (const struct ggml_tensor * tensor);
+float * ggml_get_data_f32(const struct ggml_tensor * tensor);
+
+//
+// operations on tensors with backpropagation
+//
+
+struct ggml_tensor * ggml_dup(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
+struct ggml_tensor * ggml_add(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b);
+
+struct ggml_tensor * ggml_sub(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b);
+
+struct ggml_tensor * ggml_mul(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b);
+
+struct ggml_tensor * ggml_div(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b);
+
+struct ggml_tensor * ggml_sqr(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
+struct ggml_tensor * ggml_sqrt(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
+// return scalar
+// TODO: compute sum along rows
+struct ggml_tensor * ggml_sum(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
+// mean along rows
+struct ggml_tensor * ggml_mean(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
+// if a is the same shape as b, and a is not parameter, return a
+// otherwise, return a new tensor: repeat(a) to fit in b
+struct ggml_tensor * ggml_repeat(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b);
+
+struct ggml_tensor * ggml_abs(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
+struct ggml_tensor * ggml_sgn(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
+struct ggml_tensor * ggml_neg(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
+struct ggml_tensor * ggml_step(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
+struct ggml_tensor * ggml_relu(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
+// TODO: double-check this computation is correct
+struct ggml_tensor * ggml_gelu(
+        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(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
+// A: m rows, n columns
+// B: p rows, n columns (i.e. we transpose it internally)
+// result is m columns, p rows
+struct ggml_tensor * ggml_mul_mat(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b);
+
+//
+// operations on tensors without backpropagation
+//
+
+// in-place, returns view(a)
+struct ggml_tensor * ggml_scale(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b);
+
+// a -> b, return view(b)
+struct ggml_tensor * ggml_cpy(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b);
+
+// return view(a), b specifies the new shape
+// TODO: when we start computing gradient, make a copy instead of view
+struct ggml_tensor * ggml_reshape(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b);
+
+// return view(a)
+// TODO: when we start computing gradient, make a copy instead of view
+struct ggml_tensor * ggml_reshape_2d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   ne0,
+        int                   ne1);
+
+// return view(a)
+// TODO: when we start computing gradient, make a copy instead of view
+struct ggml_tensor * ggml_reshape_3d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   ne0,
+        int                   ne1,
+        int                   ne2);
+
+// offset in bytes
+struct ggml_tensor * ggml_view_1d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   ne0,
+        size_t                offset);
+
+struct ggml_tensor * ggml_view_2d(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   ne0,
+        int                   ne1,
+        size_t                nb1, // row stride in bytes
+        size_t                offset);
+
+struct ggml_tensor * ggml_permute(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   axis0,
+        int                   axis1,
+        int                   axis2,
+        int                   axis3);
+
+// alias for ggml_permute(ctx, a, 1, 0, 2, 3)
+struct ggml_tensor * ggml_transpose(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
+struct ggml_tensor * ggml_get_rows(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b);
+
+// set elements above the diagonal to -INF
+// in-place, returns view(a)
+struct ggml_tensor * ggml_diag_mask_inf(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   n_past);
+
+// in-place, returns view(a)
+struct ggml_tensor * ggml_soft_max(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a);
+
+// rotary position embedding
+// in-place, returns view(a)
+// if mode == 1, skip n_past elements
+// TODO: avoid creating a new tensor every time
+struct ggml_tensor * ggml_rope(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        int                   n_past,
+        int                   n_dims,
+        int                   mode);
+
+// padding = 1
+// TODO: we don't support extra parameters for now
+//       that's why we are hard-coding the stride, padding, and dilation
+//       not great ..
+struct ggml_tensor * ggml_conv_1d_1s(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b);
+
+struct ggml_tensor * ggml_conv_1d_2s(
+        struct ggml_context * ctx,
+        struct ggml_tensor  * a,
+        struct ggml_tensor  * b);
+
+//
+// automatic differentiation
+//
+
+void ggml_set_param(
+        struct ggml_context * ctx,
+        struct ggml_tensor * tensor);
+
+void ggml_build_forward_expand(struct ggml_cgraph * cgraph, struct ggml_tensor * tensor);
+
+struct ggml_cgraph ggml_build_forward (struct ggml_tensor * tensor);
+struct ggml_cgraph ggml_build_backward(struct ggml_context * ctx, struct ggml_cgraph * gf, bool keep);
+
+void ggml_graph_compute(struct ggml_context * ctx, struct ggml_cgraph * cgraph);
+void ggml_graph_reset  (struct ggml_cgraph * cgraph);
+
+// print info and performance information for the graph
+void ggml_graph_print(const struct ggml_cgraph * cgraph);
+
+// dump the graph into a file using the dot format
+void ggml_graph_dump_dot(const struct ggml_cgraph * gb, const struct ggml_cgraph * gf, const char * filename);
+
+//
+// optimization
+//
+
+// optimization methods
+enum ggml_opt_type {
+    GGML_OPT_ADAM,
+    GGML_OPT_LBFGS,
+};
+
+// linesearch methods
+enum ggml_linesearch {
+    GGML_LINESEARCH_DEFAULT = 1,
+
+    GGML_LINESEARCH_BACKTRACKING_ARMIJO       = 0,
+    GGML_LINESEARCH_BACKTRACKING_WOLFE        = 1,
+    GGML_LINESEARCH_BACKTRACKING_STRONG_WOLFE = 2,
+};
+
+// optimization return values
+enum ggml_opt_result {
+    GGML_OPT_OK = 0,
+    GGML_OPT_DID_NOT_CONVERGE,
+    GGML_OPT_NO_CONTEXT,
+    GGML_OPT_INVALID_WOLFE,
+    GGML_OPT_FAIL,
+
+    GGML_LINESEARCH_FAIL = -128,
+    GGML_LINESEARCH_MINIMUM_STEP,
+    GGML_LINESEARCH_MAXIMUM_STEP,
+    GGML_LINESEARCH_MAXIMUM_ITERATIONS,
+    GGML_LINESEARCH_INVALID_PARAMETERS,
+};
+
+// optimization parameters
+//
+//   see ggml.c (ggml_opt_default_params) for default values
+//
+struct ggml_opt_params {
+    enum ggml_opt_type type;
+
+    int n_threads;
+
+    // delta-based convergence test
+    //
+    //   if past == 0 - disabled
+    //   if past > 0:
+    //     stop if |f(x) - f(x_past)| < delta * max(1, |f(x)|)
+    //
+    int past;
+    float delta;
+
+    // maximum number of iterations without improvement
+    //
+    //   if 0 - disabled
+    //   if > 0:
+    //     assume convergence if no cost improvement in this number of iterations
+    //
+    int max_no_improvement;
+
+    bool print_forward_graph;
+    bool print_backward_graph;
+
+    union {
+        // ADAM parameters
+        struct {
+            int n_iter;
+
+            float alpha; // learning rate
+            float beta1;
+            float beta2;
+            float eps;   // epsilon for numerical stability
+            float eps_f; // epsilon for convergence test
+            float eps_g; // epsilon for convergence test
+        } adam;
+
+        // LBFGS parameters
+        struct {
+            int m; // number of corrections to approximate the inv. Hessian
+            int n_iter;
+            int max_linesearch;
+
+            float eps;      // convergence tolerance
+            float ftol;     // line search tolerance
+            float wolfe;
+            float min_step;
+            float max_step;
+
+            enum ggml_linesearch linesearch;
+        } lbfgs;
+    };
+};
+
+struct ggml_opt_params ggml_opt_default_params(enum ggml_opt_type type);
+
+// optimize the function defined by the tensor f
+enum ggml_opt_result ggml_opt(
+        struct ggml_context * ctx,
+        struct ggml_opt_params params,
+        struct ggml_tensor * f);
+
+#ifdef  __cplusplus
+}
+#endif
--- a/main.cpp
+++ b/main.cpp
--- a/models/.gitignore
+++ b/models/.gitignore
@ -0,0 +1 @@
+*.bin
--- a/samples/.gitignore
+++ b/samples/.gitignore
@ -0,0 +1 @@
+*
--- a/samples/jfk.wav
+++ b/samples/jfk.wav