extern/whisper.cpp
1
0
Fork 1
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2024-11-07 08:34:37 +01:00
Code Issues Packages Projects Releases Wiki Activity

whisper : add integer quantization support (#540)

Browse Source 
* whisper : add integer quantization support

* examples : add common-ggml + prepare to add "quantize" tool

* whisper : quantization tool ready

* whisper : fix F32 support

* whisper : try to fix shared lib linkage

* wasm : update quantized models to Q5

* bench.wasm : remove "medium" button

* bench.wasm : fix custom model button

* ggml : add Q5_0 and Q5_1 WASM SIMD

* wasm : add quantized models to all WASM examples

* wasm : bump DB version number to 2

* talk-llama : update example to latest llama.cpp

* node : increase test timeout to 10s

* readme : add information for model quantization

* wasm : add links to other examples
This commit is contained in:
Georgi Gerganov 2023-04-30 18:51:57 +03:00 committed by GitHub
parent 5fd1bdd7fc
commit 794b162a46
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
41 changed files with 3183 additions and 1010 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

1
.gitignore vendored
View File

@ -23,6 +23,7 @@ build-sanitize-thread/
/talk
/talk-llama
/bench
/quantize
arm_neon.h
sync.sh

6
CMakeLists.txt
View File

@ -303,6 +303,12 @@ if (BUILD_SHARED_LIBS)
target_compile_definitions(${TARGET} PUBLIC
WHISPER_SHARED
GGML_SHARED
)
target_compile_definitions(${TARGET} PRIVATE
WHISPER_BUILD
GGML_BUILD
)
endif()

9
Makefile
View File

@ -1,4 +1,4 @@
default: main bench
default: main bench quantize
ifndef UNAME_S
UNAME_S := $(shell uname -s)
@ -243,7 +243,7 @@ libwhisper.so: ggml.o $(WHISPER_OBJ)
$(CXX) $(CXXFLAGS) -shared -o libwhisper.so ggml.o $(WHISPER_OBJ) $(LDFLAGS)
clean:
rm -f *.o main stream command talk talk-llama bench libwhisper.a libwhisper.so
rm -f *.o main stream command talk talk-llama bench quantize libwhisper.a libwhisper.so
#
# Examples
@ -251,7 +251,7 @@ clean:
CC_SDL=`sdl2-config --cflags --libs`
SRC_COMMON = examples/common.cpp
SRC_COMMON = examples/common.cpp examples/common-ggml.cpp
SRC_COMMON_SDL = examples/common-sdl.cpp
main: examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ)
@ -261,6 +261,9 @@ main: examples/main/main.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ)
bench: examples/bench/bench.cpp ggml.o $(WHISPER_OBJ)
$(CXX) $(CXXFLAGS) examples/bench/bench.cpp ggml.o $(WHISPER_OBJ) -o bench $(LDFLAGS)
quantize: examples/quantize/quantize.cpp ggml.o $(WHISPER_OBJ) $(SRC_COMMON)
$(CXX) $(CXXFLAGS) examples/quantize/quantize.cpp $(SRC_COMMON) ggml.o $(WHISPER_OBJ) -o quantize $(LDFLAGS)
stream: examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ)
$(CXX) $(CXXFLAGS) examples/stream/stream.cpp $(SRC_COMMON) $(SRC_COMMON_SDL) ggml.o $(WHISPER_OBJ) -o stream $(CC_SDL) $(LDFLAGS)

17
README.md
View File

@ -15,6 +15,7 @@ High-performance inference of [OpenAI's Whisper](https://github.com/openai/whisp
- AVX intrinsics support for x86 architectures
- VSX intrinsics support for POWER architectures
- Mixed F16 / F32 precision
- [4-bit and 5-bit integer quantization support](https://github.com/ggerganov/whisper.cpp#quantization)
- Low memory usage (Flash Attention)
- Zero memory allocations at runtime
- Runs on the CPU
@ -228,6 +229,22 @@ make large
| medium | 1.5 GB | ~1.7 GB | `fd9727b6e1217c2f614f9b698455c4ffd82463b4` |
| large | 2.9 GB | ~3.3 GB | `0f4c8e34f21cf1a914c59d8b3ce882345ad349d6` |
## Quantization
`whisper.cpp` supports integer quantization of the Whisper `ggml` models.
Quantized models require less memory and disk space and depending on the hardware can be processed more efficiently.
Here are the steps for creating and using a quantized model:
```bash
# quantize a model with Q5_0 method
make quantize
./quantize models/ggml-base.en.bin models/ggml-base.en-q5_0.bin q5_0
# run the examples as usual, specifying the quantized model file
./main -m models/ggml-base.en-q5_0.bin ./samples/gb0.wav
```
## Core ML support
On Apple Silicon devices, the Encoder inference can be executed on the Apple Neural Engine (ANE) via Core ML. This can result in significant

2
bindings/javascript/whisper.js
View File

File diff suppressed because one or more lines are too long

5
examples/CMakeLists.txt
View File

@ -21,10 +21,14 @@ set(TARGET common)
add_library(${TARGET} STATIC
common.h
common.cpp
common-ggml.h
common-ggml.cpp
)
include(DefaultTargetOptions)
target_link_libraries(${TARGET} PRIVATE whisper)
set_target_properties(${TARGET} PROPERTIES POSITION_INDEPENDENT_CODE ON)
if (WHISPER_SDL2)
@ -62,6 +66,7 @@ else()
add_subdirectory(stream)
add_subdirectory(command)
add_subdirectory(bench)
add_subdirectory(quantize)
add_subdirectory(talk)
add_subdirectory(talk-llama)
endif()

9
examples/addon.node/__test__/whisper.spec.js
View File

@ -14,9 +14,10 @@ const whisperParamsMock = {
};
describe("Run whisper.node", () => {
test("it should receive a non-empty value", async () => {
let result = await whisperAsync(whisperParamsMock);
test("it should receive a non-empty value", async () => {
let result = await whisperAsync(whisperParamsMock);
expect(result.length).toBeGreaterThan(0);
});
expect(result.length).toBeGreaterThan(0);
}, 10000);
});

6
examples/bench.wasm/CMakeLists.txt
View File

@ -31,9 +31,9 @@ endif()
set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
--bind \
-s USE_PTHREADS=1 \
-s PTHREAD_POOL_SIZE=8 \
-s INITIAL_MEMORY=1024MB \
-s TOTAL_MEMORY=1024MB \
-s PTHREAD_POOL_SIZE_STRICT=0 \
-s INITIAL_MEMORY=2000MB \
-s TOTAL_MEMORY=2000MB \
-s FORCE_FILESYSTEM=1 \
-s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
${EXTRA_FLAGS} \

74
examples/bench.wasm/index-tmpl.html
View File

@ -35,6 +35,15 @@
<br><br>
<b>More examples:</b>
<a href="https://whisper.ggerganov.com/">main</a> |
<a href="https://whisper.ggerganov.com/bench">bench</a> |
<a href="https://whisper.ggerganov.com/stream">stream</a> |
<a href="https://whisper.ggerganov.com/command">command</a> |
<a href="https://whisper.ggerganov.com/talk">talk</a> |
<br><br>
<hr>
Select the model you would like to use and click the "Bench" button.<br>
@ -44,11 +53,18 @@
<div id="model-whisper">
Whisper model: <span id="model-whisper-status"></span>
<button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
<button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
<span id="fetch-whisper-progress"></span>
<button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
<button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
<button id="fetch-whisper-small-en" onclick="loadWhisper('small.en')">small.en (466 MB)</button>
<input type="file" id="whisper-file" name="file" onchange="loadFile(event, 'whisper.bin')" />
<br><br>
Quantized models:<br><br>
<button id="fetch-whisper-tiny-en-q5_1" onclick="loadWhisper('tiny-en-q5_1')">tiny.en (Q5_1, 31 MB)</button>
<button id="fetch-whisper-base-en-q5_1" onclick="loadWhisper('base-en-q5_1')">base.en (Q5_1, 57 MB)</button>
<button id="fetch-whisper-small-en-q5_1" onclick="loadWhisper('small-en-q5_1')">small.en (Q5_1, 182 MB)</button>
<button id="fetch-whisper-medium-en-q5_0" onclick="loadWhisper('medium-en-q5_0')">medium.en (Q5_0, 515 MB)</button>
<button id="fetch-whisper-large-q5_0" onclick="loadWhisper('large-q5_0')">large (Q5_0, 1030 MB)</button>
<span id="fetch-whisper-progress"></span>
</div>
<br>
@ -160,6 +176,14 @@
document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
document.getElementById('fetch-whisper-base-en').style.display = 'none';
document.getElementById('fetch-whisper-small-en').style.display = 'none';
document.getElementById('fetch-whisper-tiny-en-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-base-en-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-small-en-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-medium-en-q5_0').style.display = 'none';
document.getElementById('fetch-whisper-large-q5_0' ).style.display = 'none';
document.getElementById('whisper-file' ).style.display = 'none';
document.getElementById('model-whisper-status' ).innerHTML = 'loaded model: ' + file.name;
}
@ -168,19 +192,42 @@
let urls = {
'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
'small.en': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en.bin',
'tiny-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
'base-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
'small-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en-q5_1.bin',
'medium-en-q5_0':'https://whisper.ggerganov.com/ggml-model-whisper-medium.en-q5_0.bin',
'large-q5_0': 'https://whisper.ggerganov.com/ggml-model-whisper-large-q5_0.bin',
};
let sizes = {
'tiny.en': 75,
'base.en': 142,
'small.en': 466,
'tiny-en-q5_1': 31,
'base-en-q5_1': 57,
'small-en-q5_1': 182,
'medium-en-q5_0': 515,
'large-q5_0': 1030,
};
let url = urls[model];
let dst = 'whisper.bin';
let size_mb = sizes[model];
document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
document.getElementById('fetch-whisper-base-en').style.display = 'none';
document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
document.getElementById('fetch-whisper-base-en').style.display = 'none';
document.getElementById('fetch-whisper-small-en').style.display = 'none';
document.getElementById('fetch-whisper-tiny-en-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-base-en-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-small-en-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-medium-en-q5_0').style.display = 'none';
document.getElementById('fetch-whisper-large-q5_0' ).style.display = 'none';
document.getElementById('whisper-file' ).style.display = 'none';
document.getElementById('model-whisper-status').innerHTML = 'loading "' + model + '" ... ';
cbProgress = function(p) {
@ -190,9 +237,18 @@
cbCancel = function() {
var el;
el = document.getElementById('fetch-whisper-tiny-en'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-base-en'); if (el) el.style.display = 'inline-block';
el = document.getElementById('model-whisper-status'); if (el) el.innerHTML = '';
el = document.getElementById('fetch-whisper-tiny-en'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-base-en'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-small-en'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-tiny-en-q5_1' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-base-en-q5_1' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-small-en-q5_1' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-medium-en-q5_0'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-large-q5_0' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('whisper-file' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('model-whisper-status'); if (el) el.innerHTML = '';
};
loadRemote(url, dst, size_mb, cbProgress, storeFS, cbCancel, printTextarea);

27
examples/command.wasm/index-tmpl.html
View File

@ -35,6 +35,15 @@
<br><br>
<b>More examples:</b>
<a href="https://whisper.ggerganov.com/">main</a> |
<a href="https://whisper.ggerganov.com/bench">bench</a> |
<a href="https://whisper.ggerganov.com/stream">stream</a> |
<a href="https://whisper.ggerganov.com/command">command</a> |
<a href="https://whisper.ggerganov.com/talk">talk</a> |
<br><br>
<hr>
Select the model you would like to use, click the "Start" button and follow the instructions.
@ -45,6 +54,10 @@
Whisper model: <span id="model-whisper-status"></span>
<button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
<button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
<br><br>
Quantized models:<br><br>
<button id="fetch-whisper-tiny-en-q5_1" onclick="loadWhisper('tiny-en-q5_1')">tiny.en (Q5_1, 31 MB)</button>
<button id="fetch-whisper-base-en-q5_1" onclick="loadWhisper('base-en-q5_1')">base.en (Q5_1, 57 MB)</button>
<span id="fetch-whisper-progress"></span>
<!--
@ -162,11 +175,17 @@
let urls = {
'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
'tiny-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
'base-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
};
let sizes = {
'tiny.en': 75,
'base.en': 142,
'tiny-en-q5_1': 31,
'base-en-q5_1': 57,
};
let url = urls[model];
@ -177,6 +196,10 @@
document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
document.getElementById('fetch-whisper-base-en').style.display = 'none';
document.getElementById('fetch-whisper-tiny-en-q5_1').style.display = 'none';
document.getElementById('fetch-whisper-base-en-q5_1').style.display = 'none';
document.getElementById('model-whisper-status').innerHTML = 'loading "' + model + '" ... ';
cbProgress = function(p) {
@ -188,6 +211,10 @@
var el;
el = document.getElementById('fetch-whisper-tiny-en'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-base-en'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-tiny-en-q5_1'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-base-en-q5_1'); if (el) el.style.display = 'inline-block';
el = document.getElementById('model-whisper-status'); if (el) el.innerHTML = '';
};

241
examples/common-ggml.cpp Normal file
View File

@ -0,0 +1,241 @@
#include "common-ggml.h"
#include <regex>
#include <map>
static const std::map<std::string, enum ggml_ftype> GGML_FTYPE_MAP = {
{"q4_0", GGML_FTYPE_MOSTLY_Q4_0},
{"q4_1", GGML_FTYPE_MOSTLY_Q4_1},
{"q4_2", GGML_FTYPE_MOSTLY_Q4_2},
{"q5_0", GGML_FTYPE_MOSTLY_Q5_0},
{"q5_1", GGML_FTYPE_MOSTLY_Q5_1},
{"q8_0", GGML_FTYPE_MOSTLY_Q8_0},
};
void ggml_print_ftypes(FILE * fp) {
for (auto it = GGML_FTYPE_MAP.begin(); it != GGML_FTYPE_MAP.end(); it++) {
fprintf(fp, " type = \"%s\" or %d\n", it->first.c_str(), it->second);
}
}
enum ggml_ftype ggml_parse_ftype(const char * str) {
enum ggml_ftype ftype;
if (str[0] == 'q') {
const auto it = GGML_FTYPE_MAP.find(str);
if (it == GGML_FTYPE_MAP.end()) {
fprintf(stderr, "%s: unknown ftype '%s'\n", __func__, str);
return GGML_FTYPE_UNKNOWN;
}
ftype = it->second;
} else {
ftype = (enum ggml_ftype) atoi(str);
}
return ftype;
}
bool ggml_common_quantize_0(
std::ifstream & finp,
std::ofstream & fout,
const ggml_ftype ftype,
const std::vector<std::string> & to_quant,
const std::vector<std::string> & to_skip) {
ggml_type qtype = GGML_TYPE_F32;
switch (ftype) {
case GGML_FTYPE_MOSTLY_Q4_0: qtype = GGML_TYPE_Q4_0; break;
case GGML_FTYPE_MOSTLY_Q4_1: qtype = GGML_TYPE_Q4_1; break;
case GGML_FTYPE_MOSTLY_Q4_2: qtype = GGML_TYPE_Q4_2; break;
case GGML_FTYPE_MOSTLY_Q5_0: qtype = GGML_TYPE_Q5_0; break;
case GGML_FTYPE_MOSTLY_Q5_1: qtype = GGML_TYPE_Q5_1; break;
case GGML_FTYPE_MOSTLY_Q8_0: qtype = GGML_TYPE_Q8_0; break;
case GGML_FTYPE_UNKNOWN:
case GGML_FTYPE_ALL_F32:
case GGML_FTYPE_MOSTLY_F16:
case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16:
{
fprintf(stderr, "%s: invalid model type %d\n", __func__, ftype);
return false;
}
};
if (!ggml_is_quantized(qtype)) {
fprintf(stderr, "%s: invalid quantization type %d (%s)\n", __func__, qtype, ggml_type_name(qtype));
return false;
}
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 ttype;
finp.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
finp.read(reinterpret_cast<char *>(&length), sizeof(length));
finp.read(reinterpret_cast<char *>(&ttype), sizeof(ttype));
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);
printf("%64s - [%5d, %5d], type = %6s ", name.data(), ne[0], ne[1], ggml_type_name((ggml_type) ttype));
bool quantize = false;
// check if we should quantize this tensor
for (const auto & s : to_quant) {
if (std::regex_match(name, std::regex(s))) {
quantize = true;
break;
}
}
// check if we should skip this tensor
for (const auto & s : to_skip) {
if (std::regex_match(name, std::regex(s))) {
quantize = false;
break;
}
}
// quantize only 2D tensors
quantize &= (n_dims == 2);
if (quantize) {
if (ttype != GGML_TYPE_F32 && ttype != GGML_TYPE_F16) {
fprintf(stderr, "%s: unsupported ttype %d (%s) for integer quantization\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
return false;
}
if (ttype == GGML_TYPE_F16) {
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));
}
ttype = qtype;
} else {
const int bpe = (ttype == 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 *>(&ttype), sizeof(ttype));
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) {
work.resize(nelements); // for quantization
size_t cur_size = 0;
std::vector<int64_t> hist_cur(1 << 4, 0);
switch ((ggml_type) ttype) {
case GGML_TYPE_Q4_0:
{
cur_size = ggml_quantize_q4_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q4_1:
{
cur_size = ggml_quantize_q4_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q4_2:
{
cur_size = ggml_quantize_q4_2(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q5_0:
{
cur_size = ggml_quantize_q5_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q5_1:
{
cur_size = ggml_quantize_q5_1(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
} break;
case GGML_TYPE_Q8_0:
{
cur_size = ggml_quantize_q8_0(data_f32.data(), work.data(), nelements, ne[0], hist_cur.data());
} break;
case GGML_TYPE_F32:
case GGML_TYPE_F16:
case GGML_TYPE_I8:
case GGML_TYPE_I16:
case GGML_TYPE_I32:
case GGML_TYPE_Q8_1:
case GGML_TYPE_COUNT:
{
fprintf(stderr, "%s: unsupported quantization type %d (%s)\n", __func__, ttype, ggml_type_name((ggml_type) ttype));
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 | ftype = %d (%s)\n", __func__, total_size_new/1024.0/1024.0, ftype, ggml_type_name(qtype));
{
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");
}
return true;
}

18
examples/common-ggml.h Normal file
View File

@ -0,0 +1,18 @@
#pragma once
#include "ggml.h"
#include <fstream>
#include <vector>
#include <string>
enum ggml_ftype ggml_parse_ftype(const char * str);
void ggml_print_ftypes(FILE * fp = stderr);
bool ggml_common_quantize_0(
std::ifstream & finp,
std::ofstream & fout,
const ggml_ftype ftype,
const std::vector<std::string> & to_quant,
const std::vector<std::string> & to_skip);

320
examples/common.cpp
View File

@ -6,13 +6,86 @@
#include "dr_wav.h"
#include <cmath>
#include <cstdint>
#include <fstream>
#include <regex>
#ifndef M_PI
#define M_PI 3.14159265358979323846
#endif
bool gpt_params_parse(int argc, char ** argv, gpt_params & params) {
for (int i = 1; i < argc; i++) {
std::string arg = argv[i];
if (arg == "-s" || arg == "--seed") {
params.seed = std::stoi(argv[++i]);
} else if (arg == "-t" || arg == "--threads") {
params.n_threads = std::stoi(argv[++i]);
} else if (arg == "-p" || arg == "--prompt") {
params.prompt = argv[++i];
} else if (arg == "-n" || arg == "--n_predict") {
params.n_predict = std::stoi(argv[++i]);
} else if (arg == "--top_k") {
params.top_k = std::stoi(argv[++i]);
} else if (arg == "--top_p") {
params.top_p = std::stof(argv[++i]);
} else if (arg == "--temp") {
params.temp = std::stof(argv[++i]);
} else if (arg == "-b" || arg == "--batch_size") {
params.n_batch = std::stoi(argv[++i]);
} else if (arg == "-m" || arg == "--model") {
params.model = argv[++i];
} else if (arg == "-h" || arg == "--help") {
gpt_print_usage(argc, argv, params);
exit(0);
} else {
fprintf(stderr, "error: unknown argument: %s\n", arg.c_str());
gpt_print_usage(argc, argv, params);
exit(0);
}
}
return true;
}
void gpt_print_usage(int /*argc*/, char ** argv, const gpt_params & params) {
fprintf(stderr, "usage: %s [options]\n", argv[0]);
fprintf(stderr, "\n");
fprintf(stderr, "options:\n");
fprintf(stderr, " -h, --help show this help message and exit\n");
fprintf(stderr, " -s SEED, --seed SEED RNG seed (default: -1)\n");
fprintf(stderr, " -t N, --threads N number of threads to use during computation (default: %d)\n", params.n_threads);
fprintf(stderr, " -p PROMPT, --prompt PROMPT\n");
fprintf(stderr, " prompt to start generation with (default: random)\n");
fprintf(stderr, " -n N, --n_predict N number of tokens to predict (default: %d)\n", params.n_predict);
fprintf(stderr, " --top_k N top-k sampling (default: %d)\n", params.top_k);
fprintf(stderr, " --top_p N top-p sampling (default: %.1f)\n", params.top_p);
fprintf(stderr, " --temp N temperature (default: %.1f)\n", params.temp);
fprintf(stderr, " -b N, --batch_size N batch size for prompt processing (default: %d)\n", params.n_batch);
fprintf(stderr, " -m FNAME, --model FNAME\n");
fprintf(stderr, " model path (default: %s)\n", params.model.c_str());
fprintf(stderr, "\n");
}
std::string gpt_random_prompt(std::mt19937 & rng) {
const int r = rng() % 10;
switch (r) {
case 0: return "So";
case 1: return "Once upon a time";
case 2: return "When";
case 3: return "The";
case 4: return "After";
case 5: return "If";
case 6: return "import";
case 7: return "He";
case 8: return "She";
case 9: return "They";
default: return "To";
}
return "The";
}
std::string trim(const std::string & s) {
std::regex e("^\\s+|\\s+$");
return std::regex_replace(s, e, "");
@ -28,6 +101,251 @@ std::string replace(const std::string & s, const std::string & from, const std::
return result;
}
std::map<std::string, int32_t> json_parse(const std::string & fname) {
std::map<std::string, int32_t> result;
// read file into string
std::string json;
{
std::ifstream ifs(fname);
if (!ifs) {
fprintf(stderr, "Failed to open %s\n", fname.c_str());
exit(1);
}
json = std::string((std::istreambuf_iterator<char>(ifs)),
(std::istreambuf_iterator<char>()));
}
if (json[0] != '{') {
return result;
}
// parse json
{
bool has_key = false;
bool in_token = false;
std::string str_key = "";
std::string str_val = "";
int n = json.size();
for (int i = 1; i < n; ++i) {
if (!in_token) {
if (json[i] == ' ') continue;
if (json[i] == '"') {
in_token = true;
continue;
}
} else {
if (json[i] == '\\' && i+1 < n) {
if (has_key == false) {
str_key += json[i];
} else {
str_val += json[i];
}
++i;
} else if (json[i] == '"') {
if (has_key == false) {
has_key = true;
++i;
while (json[i] == ' ') ++i;
++i; // :
while (json[i] == ' ') ++i;
if (json[i] != '\"') {
while (json[i] != ',' && json[i] != '}') {
str_val += json[i++];
}
has_key = false;
} else {
in_token = true;
continue;
}
} else {
has_key = false;
}
str_key = ::replace(str_key, "\\u0120", " " ); // \u0120 -> space
str_key = ::replace(str_key, "\\u010a", "\n"); // \u010a -> new line
str_key = ::replace(str_key, "\\\"", "\""); // \\\" -> "
try {
result[str_key] = std::stoi(str_val);
} catch (...) {
//fprintf(stderr, "%s: ignoring key '%s' with value '%s'\n", fname.c_str(), str_key.c_str(), str_val.c_str());
}
str_key = "";
str_val = "";
in_token = false;
continue;
}
if (has_key == false) {
str_key += json[i];
} else {
str_val += json[i];
}
}
}
}
return result;
}
std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text) {
std::vector<std::string> words;
// first split the text into words
{
std::string str = text;
std::string pat = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
std::regex re(pat);
std::smatch m;
while (std::regex_search(str, m, re)) {
for (auto x : m) {
words.push_back(x);
}
str = m.suffix();
}
}
// find the longest tokens that form the words:
std::vector<gpt_vocab::id> tokens;
for (const auto & word : words) {
if (word.size() == 0) continue;
int i = 0;
int n = word.size();
while (i < n) {
int j = n;
while (j > i) {
auto it = vocab.token_to_id.find(word.substr(i, j-i));
if (it != vocab.token_to_id.end()) {
tokens.push_back(it->second);
i = j;
break;
}
--j;
}
if (i == n) {
break;
}
if (j == i) {
auto sub = word.substr(i, 1);
if (vocab.token_to_id.find(sub) != vocab.token_to_id.end()) {
tokens.push_back(vocab.token_to_id.at(sub));
} else {
fprintf(stderr, "%s: unknown token '%s'\n", __func__, sub.data());
}
++i;
}
}
}
return tokens;
}
bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab) {
printf("%s: loading vocab from '%s'\n", __func__, fname.c_str());
vocab.token_to_id = ::json_parse(fname);
for (const auto & kv : vocab.token_to_id) {
vocab.id_to_token[kv.second] = kv.first;
}
printf("%s: vocab size = %d\n", __func__, (int) vocab.token_to_id.size());
// print the vocabulary
//for (auto kv : vocab.token_to_id) {
// printf("'%s' -> %d\n", kv.first.data(), kv.second);
//}
return true;
}
gpt_vocab::id gpt_sample_top_k_top_p(
const gpt_vocab & vocab,
const float * logits,
int top_k,
double top_p,
double temp,
std::mt19937 & rng) {
int n_logits = vocab.id_to_token.size();
std::vector<std::pair<double, gpt_vocab::id>> logits_id;
logits_id.reserve(n_logits);
{
const double scale = 1.0/temp;
for (int i = 0; i < n_logits; ++i) {
logits_id.push_back(std::make_pair(logits[i]*scale, i));
}
}
// find the top K tokens
std::partial_sort(
logits_id.begin(),
logits_id.begin() + top_k, logits_id.end(),
[](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
return a.first > b.first;
});
logits_id.resize(top_k);
double maxl = -INFINITY;
for (const auto & kv : logits_id) {
maxl = std::max(maxl, kv.first);
}
// compute probs for the top K tokens
std::vector<double> probs;
probs.reserve(logits_id.size());
double sum = 0.0;
for (const auto & kv : logits_id) {
double p = exp(kv.first - maxl);
probs.push_back(p);
sum += p;
}
// normalize the probs
for (auto & p : probs) {
p /= sum;
}
if (top_p < 1.0f) {
double cumsum = 0.0f;
for (int i = 0; i < top_k; i++) {
cumsum += probs[i];
if (cumsum >= top_p) {
top_k = i + 1;
probs.resize(top_k);
logits_id.resize(top_k);
break;
}
}
cumsum = 1.0/cumsum;
for (int i = 0; i < (int) probs.size(); i++) {
probs[i] *= cumsum;
}
}
//printf("\n");
//for (int i = 0; i < (int) probs.size(); i++) {
// printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), probs[i]);
//}
//exit(0);
std::discrete_distribution<> dist(probs.begin(), probs.end());
int idx = dist(rng);
return logits_id[idx].second;
}
bool read_wav(const std::string & fname, std::vector<float>& pcmf32, std::vector<std::vector<float>>& pcmf32s, bool stereo) {
drwav wav;
std::vector<uint8_t> wav_data; // used for pipe input from stdin

86
examples/common.h
View File

@ -1,10 +1,44 @@
// Various helper functions and utilities
#pragma once
// needs to match WHISPER_SAMPLE_RATE
#include <string>
#include <map>
#include <vector>
#include <random>
#include <thread>
#define COMMON_SAMPLE_RATE 16000
#include <vector>
#include <string>
//
// CLI argument parsing
//
struct gpt_params {
int32_t seed = -1; // RNG seed
int32_t n_threads = std::min(4, (int32_t) std::thread::hardware_concurrency());
int32_t n_predict = 200; // new tokens to predict
// sampling parameters
int32_t top_k = 40;
float top_p = 0.9f;
float temp = 0.9f;
int32_t n_batch = 8; // batch size for prompt processing
std::string model = "models/gpt-2-117M/ggml-model.bin"; // model path
std::string prompt;
};
bool gpt_params_parse(int argc, char ** argv, gpt_params & params);
void gpt_print_usage(int argc, char ** argv, const gpt_params & params);
std::string gpt_random_prompt(std::mt19937 & rng);
//
// Vocab utils
//
std::string trim(const std::string & s);
@ -13,6 +47,52 @@ std::string replace(
const std::string & from,
const std::string & to);
struct gpt_vocab {
using id = int32_t;
using token = std::string;
std::map<token, id> token_to_id;
std::map<id, token> id_to_token;
};
// poor-man's JSON parsing
std::map<std::string, int32_t> json_parse(const std::string & fname);
// split text into tokens
//
// ref: https://github.com/openai/gpt-2/blob/a74da5d99abaaba920de8131d64da2862a8f213b/src/encoder.py#L53
//
// Regex (Python):
// r"""'s|'t|'re|'ve|'m|'ll|'d| ?\p{L}+| ?\p{N}+| ?[^\s\p{L}\p{N}]+|\s+(?!\S)|\s+"""
//
// Regex (C++):
// R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)"
//
std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text);
// load the tokens from encoder.json
bool gpt_vocab_init(const std::string & fname, gpt_vocab & vocab);
// sample next token given probabilities for each embedding
//
// - consider only the top K tokens
// - from them, consider only the top tokens with cumulative probability > P
//
// TODO: not sure if this implementation is correct
// TODO: temperature is not implemented
//
gpt_vocab::id gpt_sample_top_k_top_p(
const gpt_vocab & vocab,
const float * logits,
int top_k,
double top_p,
double temp,
std::mt19937 & rng);
//
// Audio utils
//
// Read WAV audio file and store the PCM data into pcmf32
// The sample rate of the audio must be equal to COMMON_SAMPLE_RATE
// If stereo flag is set and the audio has 2 channels, the pcmf32s will contain 2 channel PCM

13
examples/helpers.js
View File

@ -145,7 +145,15 @@ function loadRemote(url, dst, size_mb, cbProgress, cbReady, cbCancel, cbPrint) {
var db = event.target.result;
var tx = db.transaction(['models'], 'readwrite');
var os = tx.objectStore('models');
var rq = os.put(data, url);
var rq = null;
try {
var rq = os.put(data, url);
} catch (e) {
cbPrint('loadRemote: failed to store "' + url + '" in the IndexedDB: \n' + e);
cbCancel();
return;
}
rq.onsuccess = function (event) {
cbPrint('loadRemote: "' + url + '" stored in the IndexedDB');
@ -180,7 +188,6 @@ function loadRemote(url, dst, size_mb, cbProgress, cbReady, cbCancel, cbPrint) {
rq.onabort = function (event) {
cbPrint('loadRemote: failed to open IndexedDB: abort');
cbCancel();
};
}

2
examples/main/main.cpp
View File

@ -496,7 +496,7 @@ bool output_json(struct whisper_context * ctx, const char * fname, const whisper
value_i("layer", whisper_model_n_text_layer(ctx), true);
end_obj();
value_i("mels", whisper_model_n_mels(ctx));
value_i("f16", whisper_model_f16(ctx), true);
value_i("ftype", whisper_model_ftype(ctx), true);
end_obj();
start_obj("params");
value_s("model", params.model.c_str());

6
examples/quantize/CMakeLists.txt Normal file
View File

@ -0,0 +1,6 @@
set(TARGET quantize)
add_executable(${TARGET} quantize.cpp)
include(DefaultTargetOptions)
target_link_libraries(${TARGET} PRIVATE common whisper ${CMAKE_THREAD_LIBS_INIT})

3
examples/quantize/README.md Normal file
View File

@ -0,0 +1,3 @@
# quantize
Tool for integer quantization of Whisper `ggml` model files

215
examples/quantize/quantize.cpp Normal file
View File

@ -0,0 +1,215 @@
#include "ggml.h"
#include "common.h"
#include "common-ggml.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
#include <regex>
// 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, ggml_ftype ftype) {
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 *) &ftype, 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;
}
}
// regexes of tensor names to not be quantized
const std::vector<std::string> to_skip = {
//"encoder.*",
"encoder.conv1.bias",
"encoder.conv2.bias",
"encoder.positional_embedding",
"decoder.positional_embedding",
};
if (!ggml_common_quantize_0(finp, fout, ftype, { ".*" }, to_skip)) {
fprintf(stderr, "%s: failed to quantize model '%s'\n", __func__, fname_inp.c_str());
return false;
}
finp.close();
fout.close();
return true;
}
int main(int argc, char ** argv) {
if (argc != 4) {
fprintf(stderr, "usage: %s model-f32.bin model-quant.bin type\n", argv[0]);
ggml_print_ftypes(stderr);
return 1;
}
// needed to initialize f16 tables
{
struct ggml_init_params params = { 0, NULL, false };
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 ggml_ftype ftype = ggml_parse_ftype(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, ggml_ftype(ftype))) {
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;
}

27
examples/stream.wasm/index-tmpl.html
View File

@ -35,6 +35,15 @@
<br><br>
<b>More examples:</b>
<a href="https://whisper.ggerganov.com/">main</a> |
<a href="https://whisper.ggerganov.com/bench">bench</a> |
<a href="https://whisper.ggerganov.com/stream">stream</a> |
<a href="https://whisper.ggerganov.com/command">command</a> |
<a href="https://whisper.ggerganov.com/talk">talk</a> |
<br><br>
<hr>
Select the model you would like to use, click the "Start" button and start speaking
@ -45,6 +54,10 @@
Whisper model: <span id="model-whisper-status"></span>
<button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
<button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
<br><br>
Quantized models:<br><br>
<button id="fetch-whisper-tiny-en-q5_1" onclick="loadWhisper('tiny-en-q5_1')">tiny.en (Q5_1, 31 MB)</button>
<button id="fetch-whisper-base-en-q5_1" onclick="loadWhisper('base-en-q5_1')">base.en (Q5_1, 57 MB)</button>
<span id="fetch-whisper-progress"></span>
<!--
@ -162,11 +175,17 @@
let urls = {
'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
'tiny-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
'base-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
};
let sizes = {
'tiny.en': 75,
'base.en': 142,
'tiny-en-q5_1': 31,
'base-en-q5_1': 57,
};
let url = urls[model];
@ -177,6 +196,10 @@
document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
document.getElementById('fetch-whisper-base-en').style.display = 'none';
document.getElementById('fetch-whisper-tiny-en-q5_1').style.display = 'none';
document.getElementById('fetch-whisper-base-en-q5_1').style.display = 'none';
document.getElementById('model-whisper-status').innerHTML = 'loading "' + model + '" ... ';
cbProgress = function(p) {
@ -188,6 +211,10 @@
var el;
el = document.getElementById('fetch-whisper-tiny-en'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-base-en'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-tiny-en-q5_1'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-base-en-q5_1'); if (el) el.style.display = 'inline-block';
el = document.getElementById('model-whisper-status'); if (el) el.innerHTML = '';
};

92
examples/talk-llama/llama_util.h → examples/talk-llama/llama-util.h Executable file → Normal file
View File

@ -21,12 +21,17 @@
#if defined(_POSIX_MAPPED_FILES)
#include <sys/mman.h>
#endif
#if defined(_POSIX_MEMLOCK_RANGE)
#include <sys/resource.h>
#endif
#endif
#endif
#if defined(_WIN32)
#define WIN32_LEAN_AND_MEAN
#define NOMINMAX
#ifndef NOMINMAX
#define NOMINMAX
#endif
#include <windows.h>
#include <io.h>
#include <stdio.h> // for _fseeki64
@ -41,8 +46,12 @@
} while (0)
#ifdef __GNUC__
#ifdef __MINGW32__
__attribute__((format(gnu_printf, 1, 2)))
#else
__attribute__((format(printf, 1, 2)))
#endif
#endif
static std::string format(const char * fmt, ...) {
va_list ap, ap2;
va_start(ap, fmt);
@ -55,7 +64,7 @@ static std::string format(const char * fmt, ...) {
va_end(ap2);
va_end(ap);
return std::string(buf.data(), size);
};
}
struct llama_file {
// use FILE * so we don't have to re-open the file to mmap
@ -162,7 +171,7 @@ struct llama_mmap {
#ifdef _POSIX_MAPPED_FILES
static constexpr bool SUPPORTED = true;
llama_mmap(struct llama_file * file) {
llama_mmap(struct llama_file * file, bool prefetch = true) {
size = file->size;
int fd = fileno(file->fp);
int flags = MAP_SHARED;
@ -170,15 +179,16 @@ struct llama_mmap {
flags |= MAP_POPULATE;
#endif
addr = mmap(NULL, file->size, PROT_READ, flags, fd, 0);
close(fd);
if (addr == MAP_FAILED) {
throw format("mmap failed: %s", strerror(errno));
}
// Advise the kernel to preload the mapped memory
if (madvise(addr, file->size, MADV_WILLNEED)) {
fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
strerror(errno));
if (prefetch) {
// Advise the kernel to preload the mapped memory
if (madvise(addr, file->size, MADV_WILLNEED)) {
fprintf(stderr, "warning: madvise(.., MADV_WILLNEED) failed: %s\n",
strerror(errno));
}
}
}
@ -188,14 +198,13 @@ struct llama_mmap {
#elif defined(_WIN32)
static constexpr bool SUPPORTED = true;
llama_mmap(struct llama_file * file) {
llama_mmap(struct llama_file * file, bool prefetch = true) {
size = file->size;
HANDLE hFile = (HANDLE) _get_osfhandle(_fileno(file->fp));
HANDLE hMapping = CreateFileMappingA(hFile, NULL, PAGE_READONLY, 0, 0, NULL);
DWORD error = GetLastError();
CloseHandle(hFile);
if (hMapping == NULL) {
throw format("CreateFileMappingA failed: %s", llama_format_win_err(error).c_str());
@ -209,14 +218,20 @@ struct llama_mmap {
throw format("MapViewOfFile failed: %s", llama_format_win_err(error).c_str());
}
// Advise the kernel to preload the mapped memory
WIN32_MEMORY_RANGE_ENTRY range;
range.VirtualAddress = addr;
range.NumberOfBytes = (SIZE_T)size;
if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
#if _WIN32_WINNT >= _WIN32_WINNT_WIN8
if (prefetch) {
// Advise the kernel to preload the mapped memory
WIN32_MEMORY_RANGE_ENTRY range;
range.VirtualAddress = addr;
range.NumberOfBytes = (SIZE_T)size;
if (!PrefetchVirtualMemory(GetCurrentProcess(), 1, &range, 0)) {
fprintf(stderr, "warning: PrefetchVirtualMemory failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
}
}
#else
#pragma message("warning: You are building for pre-Windows 8; prefetch not supported")
#endif // _WIN32_WINNT >= _WIN32_WINNT_WIN8
}
~llama_mmap() {
@ -291,8 +306,18 @@ struct llama_mlock {
if (!mlock(addr, size)) {
return true;
} else {
fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n" MLOCK_SUGGESTION,
size, this->size, std::strerror(errno));
char* errmsg = std::strerror(errno);
bool suggest = (errno == ENOMEM);
// Check if the resource limit is fine after all
struct rlimit lock_limit;
if (suggest && getrlimit(RLIMIT_MEMLOCK, &lock_limit))
suggest = false;
if (suggest && (lock_limit.rlim_max > lock_limit.rlim_cur + size))
suggest = false;
fprintf(stderr, "warning: failed to mlock %zu-byte buffer (after previously locking %zu bytes): %s\n%s",
size, this->size, errmsg, suggest ? MLOCK_SUGGESTION : "");
return false;
}
}
@ -338,8 +363,8 @@ struct llama_mlock {
// Hopefully a megabyte is enough overhead:
size_t increment = size + 1048576;
// The minimum must be <= the maximum, so we need to increase both:
min_ws_size += size;
max_ws_size += size;
min_ws_size += increment;
max_ws_size += increment;
if (!SetProcessWorkingSetSize(GetCurrentProcess(), min_ws_size, max_ws_size)) {
fprintf(stderr, "warning: SetProcessWorkingSetSize failed: %s\n",
llama_format_win_err(GetLastError()).c_str());
@ -380,4 +405,29 @@ struct llama_buffer {
delete[] addr;
}
};
#ifdef GGML_USE_CUBLAS
#include "ggml-cuda.h"
struct llama_ctx_buffer {
uint8_t * addr = NULL;
size_t size = 0;
void resize(size_t size) {
if (addr) {
ggml_cuda_host_free(addr);
}
addr = (uint8_t *) ggml_cuda_host_malloc(size);
this->size = size;
}
~llama_ctx_buffer() {
if (addr) {
ggml_cuda_host_free(addr);
}
}
};
#else
typedef llama_buffer llama_ctx_buffer;
#endif
#endif

1301
examples/talk-llama/llama.cpp
View File

File diff suppressed because it is too large Load Diff

134
examples/talk-llama/llama.h
View File

@ -39,12 +39,16 @@ extern "C" {
typedef struct llama_token_data {
llama_token id; // token id
float logit; // log-odds of the token
float p; // probability of the token
float plog; // log probability of the token
} llama_token_data;
typedef struct llama_token_data_array {
llama_token_data * data;
size_t size;
bool sorted;
} llama_token_data_array;
typedef void (*llama_progress_callback)(float progress, void *ctx);
struct llama_context_params {
@ -65,6 +69,20 @@ extern "C" {
void * progress_callback_user_data;
};
// model file types
enum llama_ftype {
LLAMA_FTYPE_ALL_F32 = 0,
LLAMA_FTYPE_MOSTLY_F16 = 1, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
LLAMA_FTYPE_MOSTLY_Q4_2 = 5, // except 1d tensors
// LLAMA_FTYPE_MOSTLY_Q4_3 (6) support has been removed
LLAMA_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors
LLAMA_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors
};
LLAMA_API struct llama_context_params llama_context_default_params();
LLAMA_API bool llama_mmap_supported();
@ -82,27 +100,46 @@ extern "C" {
// TODO: not great API - very likely to change
// Returns 0 on success
// nthread - how many threads to use. If <=0, will use std::thread::hardware_concurrency(), else the number given
LLAMA_API int llama_model_quantize(
const char * fname_inp,
const char * fname_out,
int itype);
enum llama_ftype ftype,
int nthread);
// Returns the KV cache that will contain the context for the
// ongoing prediction with the model.
LLAMA_API const uint8_t * llama_get_kv_cache(struct llama_context * ctx);
// Returns the size of the KV cache
LLAMA_API size_t llama_get_kv_cache_size(struct llama_context * ctx);
// Apply a LoRA adapter to a loaded model
// path_base_model is the path to a higher quality model to use as a base for
// the layers modified by the adapter. Can be NULL to use the current loaded model.
// The model needs to be reloaded before applying a new adapter, otherwise the adapter
// will be applied on top of the previous one
// Returns 0 on success
LLAMA_API int llama_apply_lora_from_file(
struct llama_context * ctx,
const char * path_lora,
const char * path_base_model,
int n_threads);
// Returns the number of tokens in the KV cache
LLAMA_API int llama_get_kv_cache_token_count(struct llama_context * ctx);
// Sets the KV cache containing the current context for the model
LLAMA_API void llama_set_kv_cache(
struct llama_context * ctx,
const uint8_t * kv_cache,
size_t n_size,
int n_token_count);
// Sets the current rng seed.
LLAMA_API void llama_set_rng_seed(struct llama_context * ctx, int seed);
// Returns the size in bytes of the state (rng, logits, embedding and kv_cache)
LLAMA_API size_t llama_get_state_size(struct llama_context * ctx);
// Copies the state to the specified destination address.
// Destination needs to have allocated enough memory.
// Returns the number of bytes copied
LLAMA_API size_t llama_copy_state_data(struct llama_context * ctx, uint8_t * dest);
// Set the state reading from the specified address
// Returns the number of bytes read
LLAMA_API size_t llama_set_state_data(struct llama_context * ctx, const uint8_t * src);
// Save/load session file
LLAMA_API size_t llama_load_session_file(struct llama_context * ctx, const char * path_session, llama_token * tokens_out, size_t n_token_capacity, size_t * n_token_count_out);
LLAMA_API size_t llama_save_session_file(struct llama_context * ctx, const char * path_session, const llama_token * tokens, size_t n_token_count);
// Run the llama inference to obtain the logits and probabilities for the next token.
// tokens + n_tokens is the provided batch of new tokens to process
@ -148,16 +185,52 @@ extern "C" {
// Special tokens
LLAMA_API llama_token llama_token_bos();
LLAMA_API llama_token llama_token_eos();
LLAMA_API llama_token llama_token_nl();
// TODO: improve the last_n_tokens interface ?
LLAMA_API llama_token llama_sample_top_p_top_k(
struct llama_context * ctx,
const llama_token * last_n_tokens_data,
int last_n_tokens_size,
int top_k,
float top_p,
float temp,
float repeat_penalty);
// Sampling functions
/// @details Repetition penalty described in CTRL academic paper https://arxiv.org/abs/1909.05858, with negative logit fix.
LLAMA_API void llama_sample_repetition_penalty(struct llama_context * ctx, llama_token_data_array * candidates, llama_token * last_tokens, size_t last_tokens_size, float penalty);
/// @details Frequency and presence penalties described in OpenAI API https://platform.openai.com/docs/api-reference/parameter-details.
LLAMA_API void llama_sample_frequency_and_presence_penalties(struct llama_context * ctx, llama_token_data_array * candidates, llama_token * last_tokens, size_t last_tokens_size, float alpha_frequency, float alpha_presence);
/// @details Sorts candidate tokens by their logits in descending order and calculate probabilities based on logits.
LLAMA_API void llama_sample_softmax(struct llama_context * ctx, llama_token_data_array * candidates);
/// @details Top-K sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
LLAMA_API void llama_sample_top_k(struct llama_context * ctx, llama_token_data_array * candidates, int k, size_t min_keep = 1);
/// @details Nucleus sampling described in academic paper "The Curious Case of Neural Text Degeneration" https://arxiv.org/abs/1904.09751
LLAMA_API void llama_sample_top_p(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep = 1);
/// @details Tail Free Sampling described in https://www.trentonbricken.com/Tail-Free-Sampling/.
LLAMA_API void llama_sample_tail_free(struct llama_context * ctx, llama_token_data_array * candidates, float z, size_t min_keep = 1);
/// @details Locally Typical Sampling implementation described in the paper https://arxiv.org/abs/2202.00666.
LLAMA_API void llama_sample_typical(struct llama_context * ctx, llama_token_data_array * candidates, float p, size_t min_keep = 1);
LLAMA_API void llama_sample_temperature(struct llama_context * ctx, llama_token_data_array * candidates, float temp);
/// @details Mirostat 1.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
/// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
/// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
/// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
/// @param m The number of tokens considered in the estimation of `s_hat`. This is an arbitrary value that is used to calculate `s_hat`, which in turn helps to calculate the value of `k`. In the paper, they use `m = 100`, but you can experiment with different values to see how it affects the performance of the algorithm.
/// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
LLAMA_API llama_token llama_sample_token_mirostat(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, int m, float * mu);
/// @details Mirostat 2.0 algorithm described in the paper https://arxiv.org/abs/2007.14966. Uses tokens instead of words.
/// @param candidates A vector of `llama_token_data` containing the candidate tokens, their probabilities (p), and log-odds (logit) for the current position in the generated text.
/// @param tau The target cross-entropy (or surprise) value you want to achieve for the generated text. A higher value corresponds to more surprising or less predictable text, while a lower value corresponds to less surprising or more predictable text.
/// @param eta The learning rate used to update `mu` based on the error between the target and observed surprisal of the sampled word. A larger learning rate will cause `mu` to be updated more quickly, while a smaller learning rate will result in slower updates.
/// @param mu Maximum cross-entropy. This value is initialized to be twice the target cross-entropy (`2 * tau`) and is updated in the algorithm based on the error between the target and observed surprisal.
LLAMA_API llama_token llama_sample_token_mirostat_v2(struct llama_context * ctx, llama_token_data_array * candidates, float tau, float eta, float * mu);
/// @details Selects the token with the highest probability.
LLAMA_API llama_token llama_sample_token_greedy(struct llama_context * ctx, llama_token_data_array * candidates);
/// @details Randomly selects a token from the candidates based on their probabilities.
LLAMA_API llama_token llama_sample_token(struct llama_context * ctx, llama_token_data_array * candidates);
// Performance information
LLAMA_API void llama_print_timings(struct llama_context * ctx);
@ -170,4 +243,15 @@ extern "C" {
}
#endif
// Internal API to be implemented by llama.cpp and used by tests/benchmarks only
#ifdef LLAMA_API_INTERNAL
#include <vector>
#include <string>
struct ggml_tensor;
std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
#endif
#endif // LLAMA_H

12
examples/talk-llama/llama_internal.h
View File

@ -1,12 +0,0 @@
// Internal header to be included by llama.cpp and tests/benchmarks only.
#ifndef LLAMA_INTERNAL_H
#define LLAMA_INTERNAL_H
#include <vector>
#include <string>
struct ggml_tensor;
std::vector<std::pair<std::string, struct ggml_tensor *>>& llama_internal_get_tensor_map(struct llama_context * ctx);
#endif // LLAMA_INTERNAL_H

30
examples/talk-llama/talk-llama.cpp
View File

@ -487,11 +487,37 @@ int main(int argc, char ** argv) {
{
auto logits = llama_get_logits(ctx_llama);
auto n_vocab = llama_n_vocab(ctx_llama);
logits[llama_token_eos()] = 0;
id = llama_sample_top_p_top_k(ctx_llama,
std::vector<llama_token_data> candidates;
candidates.reserve(n_vocab);
for (llama_token token_id = 0; token_id < n_vocab; token_id++) {
candidates.emplace_back(llama_token_data{token_id, logits[token_id], 0.0f});
}
llama_token_data_array candidates_p = { candidates.data(), candidates.size(), false };
// apply repeat penalty
const float nl_logit = logits[llama_token_nl()];
llama_sample_repetition_penalty(ctx_llama, &candidates_p,
embd_inp.data() + std::max(0, n_past - repeat_last_n),
repeat_last_n, top_k, top_p, temp, repeat_penalty);
repeat_last_n, repeat_penalty);
logits[llama_token_nl()] = nl_logit;
if (temp <= 0) {
// Greedy sampling
id = llama_sample_token_greedy(ctx_llama, &candidates_p);
} else {
// Temperature sampling
llama_sample_top_k(ctx_llama, &candidates_p, top_k);
llama_sample_top_p(ctx_llama, &candidates_p, top_p);
llama_sample_temperature(ctx_llama, &candidates_p, temp);
id = llama_sample_token(ctx_llama, &candidates_p);
}
}
if (id != llama_token_eos()) {

1
examples/talk.wasm/CMakeLists.txt
View File

@ -13,6 +13,7 @@ include(DefaultTargetOptions)
target_link_libraries(${TARGET} PRIVATE
whisper
common
)
unset(EXTRA_FLAGS)

336
examples/talk.wasm/gpt-2.cpp
View File

@ -1,4 +1,6 @@
#include "ggml.h"
#include "common-ggml.h"
#include "gpt-2.h"
#include <cmath>
@ -14,150 +16,6 @@
/////////////////////// GPT-2 BEGIN /////////////////////////
//
// Vocab utils
//
std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text) {
std::vector<std::string> words;
// first split the text into words
{
std::string str = text;
std::string pat = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
std::regex re(pat);
std::smatch m;
while (std::regex_search(str, m, re)) {
for (auto x : m) {
words.push_back(x);
}
str = m.suffix();
}
}
// find the longest tokens that form the words:
std::vector<gpt_vocab::id> tokens;
for (const auto & word : words) {
if (word.size() == 0) continue;
int i = 0;
int n = word.size();
while (i < n) {
int j = n;
while (j > i) {
auto it = vocab.token_to_id.find(word.substr(i, j-i));
if (it != vocab.token_to_id.end()) {
tokens.push_back(it->second);
i = j;
break;
}
--j;
}
if (i == n) {
break;
}
if (j == i) {
auto sub = word.substr(i, 1);
if (vocab.token_to_id.find(sub) != vocab.token_to_id.end()) {
tokens.push_back(vocab.token_to_id.at(sub));
} else {
fprintf(stderr, "%s: unknown token '%s'\n", __func__, sub.data());
}
++i;
}
}
}
return tokens;
}
gpt_vocab::id gpt_sample_top_k_top_p(
const gpt_vocab & vocab,
const float * logits,
int top_k,
double top_p,
double temp,
std::mt19937 & rng) {
int n_logits = vocab.id_to_token.size();
std::vector<std::pair<double, gpt_vocab::id>> logits_id;
logits_id.reserve(n_logits);
for (int i = 0; i < n_logits; i++) {
logits_id.push_back(std::make_pair(logits[i], i));
}
// find the top K tokens
std::partial_sort(
logits_id.begin(),
logits_id.begin() + top_k, logits_id.end(),
[](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
return a.first > b.first;
});
logits_id.resize(top_k);
// normalize
{
double sum = 0.0f;
for (int i = 0; i < (int)logits_id.size(); i++) {
sum += logits_id[i].first;
}
sum = 1.0/sum;
for (int i = 0; i < (int)logits_id.size(); i++) {
logits_id[i].first *= sum;
}
}
if (top_p < 1.0f) {
{
double cumsum = 0.0f;
for (int i = 0; i < top_k; i++) {
cumsum += logits_id[i].first;
if (cumsum >= top_p) {
logits_id.resize(i+1);
break;
}
}
}
// normalize again
{
double sum = 0.0f;
for (int i = 0; i < (int)logits_id.size(); i++) {
sum += logits_id[i].first;
}
sum = 1.0/sum;
for (int i = 0; i < (int)logits_id.size(); i++) {
logits_id[i].first *= sum;
}
}
}
//printf("\n");
//for (int i = 0; i < (int)logits_id.size(); i++) {
// printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), logits_id[i].first);
//}
//exit(0);
// sample from the obtained distribution
std::vector<double> probs;
probs.reserve(logits_id.size());
for (int i = 0; i < (int) logits_id.size(); i++) {
probs.push_back(logits_id[i].first);
}
std::discrete_distribution<> dist(probs.begin(), probs.end());
int idx = dist(rng);
return logits_id[idx].second;
}
// default hparams (GPT-2 117M)
struct gpt2_hparams {
int32_t n_vocab = 50257;
@ -165,7 +23,7 @@ struct gpt2_hparams {
int32_t n_embd = 768;
int32_t n_head = 12;
int32_t n_layer = 12;
int32_t f16 = 1;
int32_t ftype = 1;
};
struct gpt2_layer {
@ -187,7 +45,7 @@ struct gpt2_layer {
struct ggml_tensor * c_mlp_fc_w;
struct ggml_tensor * c_mlp_fc_b;
struct ggml_tensor * c_mlp_proj_w_trans; // transposed for efficiency
struct ggml_tensor * c_mlp_proj_w;
struct ggml_tensor * c_mlp_proj_b;
};
@ -198,8 +56,9 @@ struct gpt2_model {
struct ggml_tensor * ln_f_g;
struct ggml_tensor * ln_f_b;
struct ggml_tensor * wte; // position embedding
struct ggml_tensor * wpe; // token embedding
struct ggml_tensor * wte; // position embedding
struct ggml_tensor * wpe; // token embedding
struct ggml_tensor * lm_head; // language model head
std::vector<gpt2_layer> layers;
@ -241,14 +100,14 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
fin.read((char *) &hparams.n_embd, sizeof(hparams.n_embd));
fin.read((char *) &hparams.n_head, sizeof(hparams.n_head));
fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
fin.read((char *) &hparams.f16, sizeof(hparams.f16));
fin.read((char *) &hparams.ftype, sizeof(hparams.ftype));
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);
printf("%s: ftype = %d\n", __func__, hparams.ftype);
}
// load vocab
@ -275,9 +134,14 @@ 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_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
if (wtype == GGML_TYPE_COUNT) {
fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
__func__, fname.c_str(), model.hparams.ftype);
return false;
}
auto & ctx = model.ctx;
@ -291,32 +155,33 @@ 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_vocab*n_embd*ggml_type_sizef(wtype); // lm_head
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
@ -325,9 +190,11 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
// create the ggml context
{
struct ggml_init_params params;
params.mem_size = ctx_size;
params.mem_buffer = NULL;
struct ggml_init_params params = {
.mem_size = ctx_size,
.mem_buffer = NULL,
.no_alloc = false,
};
model.ctx = ggml_init(params);
if (!model.ctx) {
@ -350,36 +217,38 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
model.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
model.wpe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
model.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
model.wpe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
model.lm_head = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
// map by name
model.tensors["model/ln_f/g"] = model.ln_f_g;
model.tensors["model/ln_f/b"] = model.ln_f_b;
model.tensors["model/wte"] = model.wte;
model.tensors["model/wpe"] = model.wpe;
model.tensors["model/wte"] = model.wte;
model.tensors["model/wpe"] = model.wpe;
model.tensors["model/lm_head"] = model.lm_head;
for (int i = 0; i < n_layer; ++i) {
auto & layer = model.layers[i];
layer.ln_1_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.ln_1_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
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.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_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*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_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_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*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);
layer.c_mlp_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.c_mlp_proj_w = ggml_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd);
layer.c_mlp_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
// map by name
model.tensors["model/h" + std::to_string(i) + "/ln_1/g"] = layer.ln_1_g;
@ -397,7 +266,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"] = layer.c_mlp_fc_w;
model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"] = layer.c_mlp_fc_b;
model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"] = layer.c_mlp_proj_w_trans;
model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"] = layer.c_mlp_proj_w;
model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"] = layer.c_mlp_proj_b;
}
}
@ -425,14 +294,16 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
{
size_t total_size = 0;
bool has_lm_head = false;
while (true) {
int32_t n_dims;
int32_t length;
int32_t ftype;
int32_t ttype;
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));
fin.read(reinterpret_cast<char *>(&ttype), sizeof(ttype));
if (fin.eof()) {
break;
@ -461,13 +332,18 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
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]);
__func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
return false;
}
const size_t bpe = (ftype == 0) ? sizeof(float) : sizeof(ggml_fp16_t);
// for debugging
if (0) {
printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
}
if (nelements*bpe != ggml_nbytes(tensor)) {
const size_t bpe = ggml_type_size(ggml_type(ttype));
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;
@ -475,7 +351,15 @@ 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);
// GPT-2 models share the WTE tensor as the LM head
if (name == "model/wte" && has_lm_head == false) {
memcpy(model.lm_head->data, tensor->data, ggml_nbytes(tensor));
}
if (name == "model/lm_head") {
has_lm_head = true;
}
total_size += ggml_nbytes(tensor);
}
@ -493,7 +377,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
// - 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 probabilities of the next token
// - embd_w: the predicted logits for the next token
//
bool gpt2_eval(
const gpt2_model & model,
@ -512,12 +396,12 @@ bool gpt2_eval(
const int n_head = hparams.n_head;
const int n_vocab = hparams.n_vocab;
static size_t buf_size = 640u*1024*1024;
static size_t buf_size = 512u*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);
//printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
// reallocate
buf_size = buf_size_new;
@ -528,13 +412,14 @@ bool gpt2_eval(
}
}
struct ggml_init_params params;
params.mem_size = buf_size;
params.mem_buffer = buf;
struct ggml_init_params params = {
/*.mem_size =*/ buf_size,
/*.mem_buffer =*/ buf,
/*.no_alloc =*/ false,
};
struct ggml_context * ctx0 = ggml_init(params);
struct ggml_cgraph gf = { };
struct ggml_cgraph gf = {};
gf.n_threads = n_threads;
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
@ -578,7 +463,7 @@ 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,
@ -654,11 +539,13 @@ bool gpt2_eval(
// V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
// [n_past + N, 64, 12]
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);
ggml_cpy(ctx0,
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),
ggml_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd/n_head, n_head));
// KQV = transpose(V) * KQ_soft_max
// [64, N, 12]
@ -685,7 +572,7 @@ 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,
@ -722,7 +609,7 @@ 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,
@ -742,7 +629,7 @@ bool gpt2_eval(
// cur = proj_w*cur + proj_b
// [768, N]
cur = ggml_mul_mat(ctx0,
model.layers[il].c_mlp_proj_w_trans,
model.layers[il].c_mlp_proj_w,
cur);
cur = ggml_add(ctx0,
@ -769,12 +656,12 @@ bool gpt2_eval(
}
// inpL = WTE * inpL
// [ 768, 50257] - model.wte
// [ 768, 50257] - model.lm_head
// [ 768, N] - inpL
inpL = ggml_mul_mat(ctx0, model.wte, inpL);
inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
// logits -> probs
inpL = ggml_soft_max(ctx0, inpL);
//inpL = ggml_soft_max(ctx0, inpL);
// run the computation
ggml_build_forward_expand(&gf, inpL);
@ -788,7 +675,7 @@ bool gpt2_eval(
//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
// return result just for 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);
@ -825,7 +712,7 @@ Me too.
int32_t n_threads = std::min(N_THREAD, (int) std::thread::hardware_concurrency());
// sampling parameters
int32_t top_k = 40;
int32_t top_k = 5;
float top_p = 0.9f;
float temp = 1.0f;
};
@ -833,14 +720,14 @@ Me too.
struct gpt2_context * gpt2_init(const char * path_model) {
gpt2_context * ctx = new gpt2_context;
ctx->rng = std::mt19937(time(NULL));
ctx->rng = std::mt19937(time(nullptr));
// load the model
{
const int64_t t_start_us = ggml_time_us();
if (!gpt2_model_load(path_model, ctx->model, ctx->vocab)) {
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, "gpt-2.bin");
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, path_model);
delete ctx;
return nullptr;
}
@ -885,9 +772,9 @@ std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens)
std::string result;
for (int i = embd.size(); i < embd_inp.size() + n_predict; i++) {
for (int i = embd.size(); i < (int) embd_inp.size() + n_predict; i++) {
// predict
if (embd.size() > 0) {
if (!embd.empty()) {
if (!gpt2_eval(ctx->model, ctx->n_threads, n_past, embd, embd_w, mem_per_token)) {
printf("gpt-2: failed to generate text\n");
return "";
@ -914,10 +801,7 @@ std::string gpt2_gen_text(gpt2_context * ctx, const char * text, int max_tokens)
result += ctx->vocab.id_to_token[embd[0]];
// end of text token
if (embd.back() == 50256 ||
ctx->vocab.id_to_token[embd.back()] == "." ||
ctx->vocab.id_to_token[embd.back()] == "!" ||
ctx->vocab.id_to_token[embd.back()] == "?") {
if (embd.back() == 50256) {
break;
}
}

10
examples/talk.wasm/gpt-2.h
View File

@ -2,18 +2,12 @@
// TODO: Change to C-style API and move to ./examples for easy reuse.
#include "common.h"
#include <vector>
#include <map>
#include <string>
struct gpt_vocab {
using id = int32_t;
using token = std::string;
std::map<token, id> token_to_id;
std::map<id, token> id_to_token;
};
struct gpt2_context;
struct gpt2_context * gpt2_init(const char * path_model);

27
examples/talk.wasm/index-tmpl.html
View File

@ -44,6 +44,15 @@
<br><br>
<b>More examples:</b>
<a href="https://whisper.ggerganov.com/">main</a> |
<a href="https://whisper.ggerganov.com/bench">bench</a> |
<a href="https://whisper.ggerganov.com/stream">stream</a> |
<a href="https://whisper.ggerganov.com/command">command</a> |
<a href="https://whisper.ggerganov.com/talk">talk</a> |
<br><br>
<hr>
Select the models you would like to use and click the "Start" button to begin the conversation
@ -54,6 +63,10 @@
Whisper model: <span id="model-whisper-status"></span>
<button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
<button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
<br><br>
Quantized models:<br><br>
<button id="fetch-whisper-tiny-en-q5_1" onclick="loadWhisper('tiny-en-q5_1')">tiny.en (Q5_1, 31 MB)</button>
<button id="fetch-whisper-base-en-q5_1" onclick="loadWhisper('base-en-q5_1')">base.en (Q5_1, 57 MB)</button>
<span id="fetch-whisper-progress"></span>
<!--
@ -266,11 +279,17 @@
let urls = {
'tiny.en': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en.bin',
'base.en': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en.bin',
'tiny-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
'base-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
};
let sizes = {
'tiny.en': 75,
'base.en': 142,
'tiny-en-q5_1': 31,
'base-en-q5_1': 57,
};
let url = urls[model];
@ -281,6 +300,10 @@
document.getElementById('fetch-whisper-tiny-en').style.display = 'none';
document.getElementById('fetch-whisper-base-en').style.display = 'none';
document.getElementById('fetch-whisper-tiny-en-q5_1').style.display = 'none';
document.getElementById('fetch-whisper-base-en-q5_1').style.display = 'none';
document.getElementById('model-whisper-status').innerHTML = 'loading "' + model + '" ... ';
cbProgress = function(p) {
@ -292,6 +315,10 @@
var el;
el = document.getElementById('fetch-whisper-tiny-en'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-base-en'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-tiny-en-q5_1'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-base-en-q5_1'); if (el) el.style.display = 'inline-block';
el = document.getElementById('model-whisper-status'); if (el) el.innerHTML = '';
};

12
examples/talk/CMakeLists.txt
View File

@ -1,16 +1,8 @@
if (WHISPER_SDL2)
# talk
set(TARGET talk)
#add_executable(${TARGET} talk.cpp gpt-2.cpp)
#target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS})
#target_link_libraries(${TARGET} PRIVATE whisper ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
# TODO: this is temporary
# need to export ggml symbols for MSVC, but too lazy ..
add_executable(${TARGET} talk.cpp gpt-2.cpp ../common.cpp ../common-sdl.cpp ../../ggml.c ../../whisper.cpp)
add_executable(${TARGET} talk.cpp gpt-2.cpp)
target_link_libraries(${TARGET} PRIVATE common common-sdl whisper ${CMAKE_THREAD_LIBS_INIT})
include(DefaultTargetOptions)
target_include_directories(${TARGET} PRIVATE ${SDL2_INCLUDE_DIRS} ../../)
target_link_libraries(${TARGET} PRIVATE ${SDL2_LIBRARIES} ${CMAKE_THREAD_LIBS_INIT})
endif ()

325
examples/talk/gpt-2.cpp
View File

@ -1,4 +1,6 @@
#include "ggml.h"
#include "common-ggml.h"
#include "gpt-2.h"
#include <cmath>
@ -14,150 +16,6 @@
/////////////////////// GPT-2 BEGIN /////////////////////////
//
// Vocab utils
//
std::vector<gpt_vocab::id> gpt_tokenize(const gpt_vocab & vocab, const std::string & text) {
std::vector<std::string> words;
// first split the text into words
{
std::string str = text;
std::string pat = R"('s|'t|'re|'ve|'m|'ll|'d| ?[[:alpha:]]+| ?[[:digit:]]+| ?[^\s[:alpha:][:digit:]]+|\s+(?!\S)|\s+)";
std::regex re(pat);
std::smatch m;
while (std::regex_search(str, m, re)) {
for (auto x : m) {
words.push_back(x);
}
str = m.suffix();
}
}
// find the longest tokens that form the words:
std::vector<gpt_vocab::id> tokens;
for (const auto & word : words) {
if (word.empty()) continue;
int i = 0;
int n = word.size();
while (i < n) {
int j = n;
while (j > i) {
auto it = vocab.token_to_id.find(word.substr(i, j-i));
if (it != vocab.token_to_id.end()) {
tokens.push_back(it->second);
i = j;
break;
}
--j;
}
if (i == n) {
break;
}
if (j == i) {
auto sub = word.substr(i, 1);
if (vocab.token_to_id.find(sub) != vocab.token_to_id.end()) {
tokens.push_back(vocab.token_to_id.at(sub));
} else {
fprintf(stderr, "%s: unknown token '%s'\n", __func__, sub.data());
}
++i;
}
}
}
return tokens;
}
gpt_vocab::id gpt_sample_top_k_top_p(
const gpt_vocab & vocab,
const float * logits,
int top_k,
double top_p,
double /*temp*/,
std::mt19937 & rng) {
int n_logits = vocab.id_to_token.size();
std::vector<std::pair<double, gpt_vocab::id>> logits_id;
logits_id.reserve(n_logits);
for (int i = 0; i < n_logits; i++) {
logits_id.emplace_back(logits[i], i);
}
// find the top K tokens
std::partial_sort(
logits_id.begin(),
logits_id.begin() + top_k, logits_id.end(),
[](const std::pair<double, gpt_vocab::id> & a, const std::pair<double, gpt_vocab::id> & b) {
return a.first > b.first;
});
logits_id.resize(top_k);
// normalize
{
double sum = 0.0f;
for (int i = 0; i < (int)logits_id.size(); i++) {
sum += logits_id[i].first;
}
sum = 1.0/sum;
for (int i = 0; i < (int)logits_id.size(); i++) {
logits_id[i].first *= sum;
}
}
if (top_p < 1.0f) {
{
double cumsum = 0.0f;
for (int i = 0; i < top_k; i++) {
cumsum += logits_id[i].first;
if (cumsum >= top_p) {
logits_id.resize(i+1);
break;
}
}
}
// normalize again
{
double sum = 0.0f;
for (int i = 0; i < (int)logits_id.size(); i++) {
sum += logits_id[i].first;
}
sum = 1.0/sum;
for (int i = 0; i < (int)logits_id.size(); i++) {
logits_id[i].first *= sum;
}
}
}
//printf("\n");
//for (int i = 0; i < (int) logits_id.size(); i++) {
// printf("%d: '%s' %f\n", i, vocab.id_to_token.at(logits_id[i].second).c_str(), logits_id[i].first);
//}
//exit(0);
// sample from the obtained distribution
std::vector<double> probs;
probs.reserve(logits_id.size());
for (int i = 0; i < (int) logits_id.size(); i++) {
probs.push_back(logits_id[i].first);
}
std::discrete_distribution<> dist(probs.begin(), probs.end());
int idx = dist(rng);
return logits_id[idx].second;
}
// default hparams (GPT-2 117M)
struct gpt2_hparams {
int32_t n_vocab = 50257;
@ -165,7 +23,7 @@ struct gpt2_hparams {
int32_t n_embd = 768;
int32_t n_head = 12;
int32_t n_layer = 12;
int32_t f16 = 1;
int32_t ftype = 1;
};
struct gpt2_layer {
@ -187,7 +45,7 @@ struct gpt2_layer {
struct ggml_tensor * c_mlp_fc_w;
struct ggml_tensor * c_mlp_fc_b;
struct ggml_tensor * c_mlp_proj_w_trans; // transposed for efficiency
struct ggml_tensor * c_mlp_proj_w;
struct ggml_tensor * c_mlp_proj_b;
};
@ -198,8 +56,9 @@ struct gpt2_model {
struct ggml_tensor * ln_f_g;
struct ggml_tensor * ln_f_b;
struct ggml_tensor * wte; // position embedding
struct ggml_tensor * wpe; // token embedding
struct ggml_tensor * wte; // position embedding
struct ggml_tensor * wpe; // token embedding
struct ggml_tensor * lm_head; // language model head
std::vector<gpt2_layer> layers;
@ -241,14 +100,14 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
fin.read((char *) &hparams.n_embd, sizeof(hparams.n_embd));
fin.read((char *) &hparams.n_head, sizeof(hparams.n_head));
fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
fin.read((char *) &hparams.f16, sizeof(hparams.f16));
fin.read((char *) &hparams.ftype, sizeof(hparams.ftype));
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);
printf("%s: ftype = %d\n", __func__, hparams.ftype);
}
// load vocab
@ -268,16 +127,21 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
fin.read((char *) &len, sizeof(len));
word.resize(len);
fin.read((char *) &word[0], len);
fin.read((char *) word.data(), len);
vocab.token_to_id[word] = i;
vocab.id_to_token[i] = word;
}
}
// 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_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
if (wtype == GGML_TYPE_COUNT) {
fprintf(stderr, "%s: invalid model file '%s' (bad ftype value %d)\n",
__func__, fname.c_str(), model.hparams.ftype);
return false;
}
auto & ctx = model.ctx;
@ -291,32 +155,33 @@ 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_vocab*n_embd*ggml_type_sizef(wtype); // lm_head
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
@ -325,9 +190,11 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
// create the ggml context
{
struct ggml_init_params params;
params.mem_size = ctx_size;
params.mem_buffer = nullptr;
struct ggml_init_params params = {
.mem_size = ctx_size,
.mem_buffer = NULL,
.no_alloc = false,
};
model.ctx = ggml_init(params);
if (!model.ctx) {
@ -350,36 +217,38 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
model.ln_f_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
model.ln_f_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
model.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
model.wpe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
model.wte = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
model.wpe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_embd, n_ctx);
model.lm_head = ggml_new_tensor_2d(ctx, wtype, n_embd, n_vocab);
// map by name
model.tensors["model/ln_f/g"] = model.ln_f_g;
model.tensors["model/ln_f/b"] = model.ln_f_b;
model.tensors["model/wte"] = model.wte;
model.tensors["model/wpe"] = model.wpe;
model.tensors["model/wte"] = model.wte;
model.tensors["model/wpe"] = model.wpe;
model.tensors["model/lm_head"] = model.lm_head;
for (int i = 0; i < n_layer; ++i) {
auto & layer = model.layers[i];
layer.ln_1_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.ln_1_g = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
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.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_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 3*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_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_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*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);
layer.c_mlp_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
layer.c_mlp_proj_w = ggml_new_tensor_2d(ctx, wtype, 4*n_embd, n_embd);
layer.c_mlp_proj_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_embd);
// map by name
model.tensors["model/h" + std::to_string(i) + "/ln_1/g"] = layer.ln_1_g;
@ -397,7 +266,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/w"] = layer.c_mlp_fc_w;
model.tensors["model/h" + std::to_string(i) + "/mlp/c_fc/b"] = layer.c_mlp_fc_b;
model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"] = layer.c_mlp_proj_w_trans;
model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/w"] = layer.c_mlp_proj_w;
model.tensors["model/h" + std::to_string(i) + "/mlp/c_proj/b"] = layer.c_mlp_proj_b;
}
}
@ -425,14 +294,16 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
{
size_t total_size = 0;
bool has_lm_head = false;
while (true) {
int32_t n_dims;
int32_t length;
int32_t ftype;
int32_t ttype;
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));
fin.read(reinterpret_cast<char *>(&ttype), sizeof(ttype));
if (fin.eof()) {
break;
@ -448,7 +319,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
std::string name(length, 0);
fin.read(&name[0], length);
if (model.tensors.find(name) == model.tensors.end()) {
if (model.tensors.find(name.data()) == model.tensors.end()) {
fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
return false;
}
@ -461,13 +332,18 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
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]);
__func__, name.data(), (int) tensor->ne[0], (int) tensor->ne[1], ne[0], ne[1]);
return false;
}
const size_t bpe = (ftype == 0) ? sizeof(float) : sizeof(ggml_fp16_t);
// for debugging
if (0) {
printf("%24s - [%5d, %5d], type = %6s, %6.2f MB, %9zu bytes\n", name.data(), ne[0], ne[1], ggml_type_name(ggml_type(ttype)), ggml_nbytes(tensor)/1024.0/1024.0, ggml_nbytes(tensor));
}
if (nelements*bpe != ggml_nbytes(tensor)) {
const size_t bpe = ggml_type_size(ggml_type(ttype));
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;
@ -475,7 +351,15 @@ 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);
// GPT-2 models share the WTE tensor as the LM head
if (name == "model/wte" && has_lm_head == false) {
memcpy(model.lm_head->data, tensor->data, ggml_nbytes(tensor));
}
if (name == "model/lm_head") {
has_lm_head = true;
}
total_size += ggml_nbytes(tensor);
}
@ -493,7 +377,7 @@ bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab &
// - 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 probabilities of the next token
// - embd_w: the predicted logits for the next token
//
bool gpt2_eval(
const gpt2_model & model,
@ -512,12 +396,12 @@ bool gpt2_eval(
const int n_head = hparams.n_head;
const int n_vocab = hparams.n_vocab;
static size_t buf_size = 5640ull*1024*1024;
static size_t buf_size = 512u*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);
//printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
// reallocate
buf_size = buf_size_new;
@ -528,13 +412,14 @@ bool gpt2_eval(
}
}
struct ggml_init_params params;
params.mem_size = buf_size;
params.mem_buffer = buf;
struct ggml_init_params params = {
/*.mem_size =*/ buf_size,
/*.mem_buffer =*/ buf,
/*.no_alloc =*/ false,
};
struct ggml_context * ctx0 = ggml_init(params);
struct ggml_cgraph gf = { };
struct ggml_cgraph gf = {};
gf.n_threads = n_threads;
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
@ -578,7 +463,7 @@ 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,
@ -654,11 +539,13 @@ bool gpt2_eval(
// V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
// [n_past + N, 64, 12]
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);
ggml_cpy(ctx0,
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),
ggml_new_tensor_3d(ctx0, model.memory_v->type, n_past + N, n_embd/n_head, n_head));
// KQV = transpose(V) * KQ_soft_max
// [64, N, 12]
@ -685,7 +572,7 @@ 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,
@ -722,7 +609,7 @@ 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,
@ -742,7 +629,7 @@ bool gpt2_eval(
// cur = proj_w*cur + proj_b
// [768, N]
cur = ggml_mul_mat(ctx0,
model.layers[il].c_mlp_proj_w_trans,
model.layers[il].c_mlp_proj_w,
cur);
cur = ggml_add(ctx0,
@ -769,12 +656,12 @@ bool gpt2_eval(
}
// inpL = WTE * inpL
// [ 768, 50257] - model.wte
// [ 768, 50257] - model.lm_head
// [ 768, N] - inpL
inpL = ggml_mul_mat(ctx0, model.wte, inpL);
inpL = ggml_mul_mat(ctx0, model.lm_head, inpL);
// logits -> probs
inpL = ggml_soft_max(ctx0, inpL);
//inpL = ggml_soft_max(ctx0, inpL);
// run the computation
ggml_build_forward_expand(&gf, inpL);
@ -788,7 +675,7 @@ bool gpt2_eval(
//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
// return result just for 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);

10
examples/talk/gpt-2.h
View File

@ -2,18 +2,12 @@
// TODO: Change to C-style API and move to ./examples for easy reuse.
#include "common.h"
#include <vector>
#include <map>
#include <string>
struct gpt_vocab {
using id = int32_t;
using token = std::string;
std::map<token, id> token_to_id;
std::map<id, token> id_to_token;
};
struct gpt2_context;
struct gpt2_context * gpt2_init(const char * path_model);

6
examples/whisper.wasm/CMakeLists.txt
View File

@ -31,9 +31,9 @@ endif()
set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
--bind \
-s USE_PTHREADS=1 \
-s PTHREAD_POOL_SIZE=8 \
-s INITIAL_MEMORY=1500MB \
-s TOTAL_MEMORY=1500MB \
-s PTHREAD_POOL_SIZE_STRICT=0 \
-s INITIAL_MEMORY=2000MB \
-s TOTAL_MEMORY=2000MB \
-s FORCE_FILESYSTEM=1 \
-s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
${EXTRA_FLAGS} \

10
examples/whisper.wasm/emscripten.cpp
View File

@ -10,6 +10,12 @@ std::thread g_worker;
std::vector<struct whisper_context *> g_contexts(4, nullptr);
static inline int mpow2(int n) {
int p = 1;
while (p <= n) p *= 2;
return p/2;
}
EMSCRIPTEN_BINDINGS(whisper) {
emscripten::function("init", emscripten::optional_override([](const std::string & path_model) {
if (g_worker.joinable()) {
@ -43,7 +49,7 @@ EMSCRIPTEN_BINDINGS(whisper) {
}
}));
emscripten::function("full_default", emscripten::optional_override([](size_t index, const emscripten::val & audio, const std::string & lang, bool translate) {
emscripten::function("full_default", emscripten::optional_override([](size_t index, const emscripten::val & audio, const std::string & lang, int nthreads, bool translate) {
if (g_worker.joinable()) {
g_worker.join();
}
@ -66,7 +72,7 @@ EMSCRIPTEN_BINDINGS(whisper) {
params.print_special = false;
params.translate = translate;
params.language = whisper_is_multilingual(g_contexts[index]) ? lang.c_str() : "en";
params.n_threads = std::min(8, (int) std::thread::hardware_concurrency());
params.n_threads = std::min(nthreads, std::min(16, mpow2(std::thread::hardware_concurrency())));
params.offset_ms = 0;
std::vector<float> pcmf32;

131
examples/whisper.wasm/index-tmpl.html
View File

@ -40,21 +40,42 @@
Note that the computation is quite heavy and may take a few seconds to complete.<br>
The transcription results will be displayed in the text area below.<br><br>
<b>Important: your browser must support WASM SIMD instructions for this to work.</b>
<b>Important:</b>
<ul>
<li>your browser must support WASM SIMD instructions for this to work</li>
<li>Firefox cannot load files larger than 256 MB - use Chrome instead</li>
</ul>
<br><br><hr>
<b>More examples:</b>
<a href="https://whisper.ggerganov.com/">main</a> |
<a href="https://whisper.ggerganov.com/bench">bench</a> |
<a href="https://whisper.ggerganov.com/stream">stream</a> |
<a href="https://whisper.ggerganov.com/command">command</a> |
<a href="https://whisper.ggerganov.com/talk">talk</a> |
<hr>
<div id="model">
Whisper model: <span id="model-whisper-status"></span>
Whisper models: <span id="model-whisper-status"></span><br><br>
<button id="fetch-whisper-tiny-en" onclick="loadWhisper('tiny.en')">tiny.en (75 MB)</button>
<button id="fetch-whisper-tiny" onclick="loadWhisper('tiny')">tiny (75 MB)</button>
<button id="fetch-whisper-base-en" onclick="loadWhisper('base.en')">base.en (142 MB)</button>
<button id="fetch-whisper-base" onclick="loadWhisper('base')">base (142 MB)</button>
<button id="fetch-whisper-small-en" onclick="loadWhisper('small.en')">small.en (466 MB)</button>
<button id="fetch-whisper-small" onclick="loadWhisper('small')">small (466 MB)</button>
<span id="fetch-whisper-progress"></span>
<input type="file" id="whisper-file" name="file" onchange="loadFile(event, 'whisper.bin')" />
<br><br>
Quantized models:<br><br>
<button id="fetch-whisper-tiny-en-q5_1" onclick="loadWhisper('tiny-en-q5_1')">tiny.en (Q5_1, 31 MB)</button>
<button id="fetch-whisper-tiny-q5_1" onclick="loadWhisper('tiny-q5_1')">tiny (Q5_1, 31 MB)</button>
<button id="fetch-whisper-base-en-q5_1" onclick="loadWhisper('base-en-q5_1')">base.en (Q5_1, 57 MB)</button>
<button id="fetch-whisper-base-q5_1" onclick="loadWhisper('base-q5_1')">base (Q5_1, 57 MB)</button>
<button id="fetch-whisper-small-en-q5_1" onclick="loadWhisper('small-en-q5_1')">small.en (Q5_1, 182 MB)</button>
<button id="fetch-whisper-small-q5_1" onclick="loadWhisper('small-q5_1')">small (Q5_1, 182 MB)</button><br>
<button id="fetch-whisper-medium-en-q5_0" onclick="loadWhisper('medium-en-q5_0')">medium.en (Q5_0, 515 MB)</button>
<button id="fetch-whisper-medium-q5_0" onclick="loadWhisper('medium-q5_0')">medium (Q5_0, 515 MB)</button>
<button id="fetch-whisper-large-q5_0" onclick="loadWhisper('large-q5_0')">large (Q5_0, 1030 MB)</button>
<span id="fetch-whisper-progress"></span>
</div>
<br>
@ -161,6 +182,12 @@
<option value="yi">Yiddish</option>
</select>
</td>
<!-- Slider to select number of threads between 1 and 16 -->
<td>
Threads:
<input type="range" id="threads" name="threads" min="1" max="16" value="8" onchange="changeThreads(this.value)" />
<span id="threads-value">8</span>
</td>
<td>
<button onclick="onProcess(false);">Transcribe</button>
</td>
@ -263,11 +290,13 @@
Module.FS_createDataFile("/", fname, buf, true, true);
model_whisper = fname;
//model_whisper = fname;
document.getElementById('model-whisper-status').innerHTML = 'loaded "' + model_whisper + '"!';
printTextarea('storeFS: stored model: ' + fname + ' size: ' + buf.length);
document.getElementById('model').innerHTML = 'Model fetched: ' + model_whisper;
}
function loadFile(event, fname) {
@ -292,6 +321,17 @@
document.getElementById('fetch-whisper-tiny' ).style.display = 'none';
document.getElementById('fetch-whisper-base' ).style.display = 'none';
document.getElementById('fetch-whisper-small' ).style.display = 'none';
document.getElementById('fetch-whisper-tiny-en-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-tiny-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-base-en-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-base-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-small-en-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-small-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-medium-en-q5_0').style.display = 'none';
document.getElementById('fetch-whisper-medium-q5_0' ).style.display = 'none';
document.getElementById('fetch-whisper-large-q5_0' ).style.display = 'none';
document.getElementById('whisper-file' ).style.display = 'none';
document.getElementById('model-whisper-status' ).innerHTML = 'loaded model: ' + file.name;
}
@ -304,6 +344,16 @@
'base': 'https://whisper.ggerganov.com/ggml-model-whisper-base.bin',
'small.en': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en.bin',
'small': 'https://whisper.ggerganov.com/ggml-model-whisper-small.bin',
'tiny-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny.en-q5_1.bin',
'tiny-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-tiny-q5_1.bin',
'base-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-base.en-q5_1.bin',
'base-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-base-q5_1.bin',
'small-en-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-small.en-q5_1.bin',
'small-q5_1': 'https://whisper.ggerganov.com/ggml-model-whisper-small-q5_1.bin',
'medium-en-q5_0':'https://whisper.ggerganov.com/ggml-model-whisper-medium.en-q5_0.bin',
'medium-q5_0': 'https://whisper.ggerganov.com/ggml-model-whisper-medium-q5_0.bin',
'large-q5_0': 'https://whisper.ggerganov.com/ggml-model-whisper-large-q5_0.bin',
};
let sizes = {
@ -313,6 +363,16 @@
'base': 142,
'small.en': 466,
'small': 466,
'tiny-en-q5_1': 31,
'tiny-q5_1': 31,
'base-en-q5_1': 57,
'base-q5_1': 57,
'small-en-q5_1': 182,
'small-q5_1': 182,
'medium-en-q5_0': 515,
'medium-q5_0': 515,
'large-q5_0': 1030,
};
let url = urls[model];
@ -327,8 +387,19 @@
document.getElementById('fetch-whisper-tiny' ).style.display = 'none';
document.getElementById('fetch-whisper-base' ).style.display = 'none';
document.getElementById('fetch-whisper-small' ).style.display = 'none';
document.getElementById('whisper-file' ).style.display = 'none';
document.getElementById('model-whisper-status' ).innerHTML = 'loading model: ' + model;
document.getElementById('fetch-whisper-tiny-en-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-tiny-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-base-en-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-base-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-small-en-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-small-q5_1' ).style.display = 'none';
document.getElementById('fetch-whisper-medium-en-q5_0').style.display = 'none';
document.getElementById('fetch-whisper-medium-q5_0' ).style.display = 'none';
document.getElementById('fetch-whisper-large-q5_0' ).style.display = 'none';
document.getElementById('whisper-file' ).style.display = 'none';
document.getElementById('model-whisper-status').innerHTML = 'loading model: ' + model;
cbProgress = function(p) {
let el = document.getElementById('fetch-whisper-progress');
@ -337,14 +408,26 @@
cbCancel = function() {
var el;
el = document.getElementById('fetch-whisper-tiny-en' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-base-en' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-small-en'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-tiny' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-base' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-small' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('whisper-file' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('model-whisper-status' ); if (el) el.innerHTML = '';
el = document.getElementById('fetch-whisper-tiny-en-q5_1' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-tiny-q5_1' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-base-en-q5_1' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-base-q5_1' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-small-en-q5_1' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-small-q5_1' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-medium-en-q5_0'); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-medium-q5_0' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('fetch-whisper-large-q5_0' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('whisper-file' ); if (el) el.style.display = 'inline-block';
el = document.getElementById('model-whisper-status'); if (el) el.innerHTML = '';
};
loadRemote(url, dst, size_mb, cbProgress, storeFS, cbCancel, printTextarea);
@ -354,7 +437,8 @@
// audio file
//
const kMaxAudio_s = 120;
const kMaxAudio_s = 30*60;
const kMaxRecording_s = 2*60;
const kSampleRate = 16000;
window.AudioContext = window.AudioContext || window.webkitAudioContext;
@ -423,7 +507,7 @@
doRecording = false;
}
// record up to kMaxAudio_s seconds of audio from the microphone
// record up to kMaxRecording_s seconds of audio from the microphone
// check if doRecording is false every 1000 ms and stop recording if so
// update progress information
function startRecording() {
@ -479,9 +563,9 @@
printTextarea('js: audio recorded, size: ' + audio.length);
// truncate to first 30 seconds
if (audio.length > kMaxAudio_s*kSampleRate) {
audio = audio.slice(0, kMaxAudio_s*kSampleRate);
printTextarea('js: truncated audio to first ' + kMaxAudio_s + ' seconds');
if (audio.length > kMaxRecording_s*kSampleRate) {
audio = audio.slice(0, kMaxRecording_s*kSampleRate);
printTextarea('js: truncated audio to first ' + kMaxRecording_s + ' seconds');
}
setAudio(audio);
});
@ -509,24 +593,31 @@
});
}
document.getElementById('progress-bar').style.width = (100*(Date.now() - startTime)/1000/kMaxAudio_s) + '%';
document.getElementById('progress-text').innerHTML = (100*(Date.now() - startTime)/1000/kMaxAudio_s).toFixed(0) + '%';
document.getElementById('progress-bar').style.width = (100*(Date.now() - startTime)/1000/kMaxRecording_s) + '%';
document.getElementById('progress-text').innerHTML = (100*(Date.now() - startTime)/1000/kMaxRecording_s).toFixed(0) + '%';
}, 1000);
printTextarea('js: recording ...');
setTimeout(function() {
if (doRecording) {
printTextarea('js: recording stopped after ' + kMaxAudio_s + ' seconds');
printTextarea('js: recording stopped after ' + kMaxRecording_s + ' seconds');
stopRecording();
}
}, kMaxAudio_s*1000);
}, kMaxRecording_s*1000);
}
//
// transcribe
//
var nthreads = 8;
function changeThreads(value) {
nthreads = value;
document.getElementById('threads-value').innerHTML = nthreads;
}
function onProcess(translate) {
if (!instance) {
instance = Module.init('whisper.bin');
@ -553,7 +644,7 @@
printTextarea('');
setTimeout(function() {
var ret = Module.full_default(instance, audio, document.getElementById('language').value, translate);
var ret = Module.full_default(instance, audio, document.getElementById('language').value, nthreads, translate);
console.log('js: full_default returned: ' + ret);
if (ret) {
printTextarea("js: whisper returned: " + ret);

45
extra/quantize-all.sh Executable file
View File

@ -0,0 +1,45 @@
#!/bin/bash
printf "Usage: $0 <upload>"
if [ $# -ne 1 ]; then
printf "\nError: Invalid number of arguments\n"
exit 1
fi
qtype0="q5_0"
qtype1="q5_1"
upload="$1"
cd `dirname $0`
cd ../
./quantize ./models/ggml-tiny.en.bin ./models/ggml-tiny.en-${qtype1}.bin ${qtype1}
./quantize ./models/ggml-tiny.bin ./models/ggml-tiny-${qtype1}.bin ${qtype1}
./quantize ./models/ggml-base.en.bin ./models/ggml-base.en-${qtype1}.bin ${qtype1}
./quantize ./models/ggml-base.bin ./models/ggml-base-${qtype1}.bin ${qtype1}
./quantize ./models/ggml-small.en.bin ./models/ggml-small.en-${qtype1}.bin ${qtype1}
./quantize ./models/ggml-small.bin ./models/ggml-small-${qtype1}.bin ${qtype1}
./quantize ./models/ggml-medium.en.bin ./models/ggml-medium.en-${qtype0}.bin ${qtype0}
./quantize ./models/ggml-medium.bin ./models/ggml-medium-${qtype0}.bin ${qtype0}
./quantize ./models/ggml-large.bin ./models/ggml-large-${qtype0}.bin ${qtype0}
if [ "$upload" == "1" ]; then
scp ./models/ggml-tiny.en-${qtype1}.bin root@linode0:/mnt/Data/ggml/ggml-model-whisper-tiny.en-${qtype1}.bin
scp ./models/ggml-tiny-${qtype1}.bin root@linode0:/mnt/Data/ggml/ggml-model-whisper-tiny-${qtype1}.bin
scp ./models/ggml-base.en-${qtype1}.bin root@linode0:/mnt/Data/ggml/ggml-model-whisper-base.en-${qtype1}.bin
scp ./models/ggml-base-${qtype1}.bin root@linode0:/mnt/Data/ggml/ggml-model-whisper-base-${qtype1}.bin
scp ./models/ggml-small.en-${qtype1}.bin root@linode0:/mnt/Data/ggml/ggml-model-whisper-small.en-${qtype1}.bin
scp ./models/ggml-small-${qtype1}.bin root@linode0:/mnt/Data/ggml/ggml-model-whisper-small-${qtype1}.bin
scp ./models/ggml-medium.en-${qtype0}.bin root@linode0:/mnt/Data/ggml/ggml-model-whisper-medium.en-${qtype0}.bin
scp ./models/ggml-medium-${qtype0}.bin root@linode0:/mnt/Data/ggml/ggml-model-whisper-medium-${qtype0}.bin
scp ./models/ggml-large-${qtype0}.bin root@linode0:/mnt/Data/ggml/ggml-model-whisper-large-${qtype0}.bin
fi

12
extra/sync-ggml.sh
View File

@ -1,6 +1,10 @@
#!/bin/bash
cp -rpv ../ggml/src/ggml.c ./ggml.c
cp -rpv ../ggml/src/ggml-cuda.cu ./ggml-cuda.cu
cp -rpv ../ggml/src/ggml-cuda.h ./ggml-cuda.h
cp -rpv ../ggml/include/ggml/ggml.h ./ggml.h
cp -rpv ../ggml/src/ggml.c ./ggml.c
cp -rpv ../ggml/src/ggml-cuda.cu ./ggml-cuda.cu
cp -rpv ../ggml/src/ggml-cuda.h ./ggml-cuda.h
cp -rpv ../ggml/include/ggml/ggml.h ./ggml.h
cp -rpv ../ggml/examples/common.h ./examples/common.h
cp -rpv ../ggml/examples/common.cpp ./examples/common.cpp
cp -rpv ../ggml/examples/common-ggml.h ./examples/common-ggml.h
cp -rpv ../ggml/examples/common-ggml.cpp ./examples/common-ggml.cpp

162
ggml.c
View File

@ -330,7 +330,7 @@ static ggml_fp16_t table_exp_f16[1 << 16];
// precomputed f32 table for f16 (256 KB)
static float table_f32_f16[1 << 16];
#if defined(__ARM_NEON)
#if defined(__ARM_NEON) || defined(__wasm_simd128__)
#define B1(c,s,n) 0x ## n ## c , 0x ## n ## s
#define B2(c,s,n) B1(c,s,n ## c), B1(c,s,n ## s)
#define B3(c,s,n) B2(c,s,n ## c), B2(c,s,n ## s)
@ -3180,6 +3180,72 @@ static void ggml_vec_dot_q5_0_q8_0(const int n, float * restrict s, const void *
}
*s = vaddvq_f32(sumv);
#elif defined(__wasm_simd128__)
v128_t sumv = wasm_f32x4_splat(0.0f);
uint64_t tmp[4];
for (int i = 0; i < nb; ++i) {
const block_q5_0 * restrict x0 = &x[i];
const block_q8_0 * restrict y0 = &y[i];
const v128_t m4b = wasm_i8x16_splat(0x0F);
const v128_t s16b = wasm_i8x16_splat(0x10);
// extract the 5th bit
uint32_t qh;
memcpy(&qh, x0->qh, sizeof(qh));
tmp[0] = table_b2b_u[(qh >> 0) & 0xFF];
tmp[1] = table_b2b_u[(qh >> 8) & 0xFF];
tmp[2] = table_b2b_u[(qh >> 16) & 0xFF];
tmp[3] = table_b2b_u[(qh >> 24) ];
const v128_t qhl = wasm_v128_load(tmp + 0);
const v128_t qhh = wasm_v128_load(tmp + 2);
const v128_t v0 = wasm_v128_load(x0->qs);
// 4-bit -> 8-bit
const v128_t v0l = wasm_v128_and (v0, m4b);
const v128_t v0h = wasm_u8x16_shr(v0, 4);
// interleave
const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h, 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23);
const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h, 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31);
// add high bit and sub 16
const v128_t v0lf = wasm_i8x16_sub(wasm_v128_or(v0lz, qhl), s16b);
const v128_t v0hf = wasm_i8x16_sub(wasm_v128_or(v0hz, qhh), s16b);
// load y
const v128_t v1l = wasm_v128_load(y0->qs);
const v128_t v1h = wasm_v128_load(y0->qs + 16);
// int8x16 -> int16x8
const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
const float x0d = GGML_FP16_TO_FP32(x0->d);
// dot product
sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
wasm_i32x4_add(
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d)));
}
*s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3);
#elif defined(__AVX2__)
// Initialize accumulator with zeros
__m256 acc = _mm256_setzero_ps();
@ -3311,6 +3377,77 @@ static void ggml_vec_dot_q5_1_q8_1(const int n, float * restrict s, const void *
}
*s = vaddvq_f32(sumv) + summs;
#elif defined(__wasm_simd128__)
v128_t sumv = wasm_f32x4_splat(0.0f);
float summs = 0.0f;
uint64_t tmp[4];
for (int i = 0; i < nb; ++i) {
const block_q5_1 * restrict x0 = &x[i];
const block_q8_1 * restrict y0 = &y[i];
summs += GGML_FP16_TO_FP32(x0->m) * (y0->s0 + y0->s1);
const v128_t m4b = wasm_i8x16_splat(0x0F);
// extract the 5th bit
uint32_t qh;
memcpy(&qh, x0->qh, sizeof(qh));
tmp[0] = table_b2b_u[(qh >> 0) & 0xFF];
tmp[1] = table_b2b_u[(qh >> 8) & 0xFF];
tmp[2] = table_b2b_u[(qh >> 16) & 0xFF];
tmp[3] = table_b2b_u[(qh >> 24) ];
const v128_t qhl = wasm_v128_load(tmp + 0);
const v128_t qhh = wasm_v128_load(tmp + 2);
const v128_t v0 = wasm_v128_load(x0->qs);
// 4-bit -> 8-bit
const v128_t v0l = wasm_v128_and (v0, m4b);
const v128_t v0h = wasm_u8x16_shr(v0, 4);
static bool x = true;
// interleave
const v128_t v0lz = wasm_v8x16_shuffle(v0l, v0h, 0, 16, 1, 17, 2, 18, 3, 19, 4, 20, 5, 21, 6, 22, 7, 23);
const v128_t v0hz = wasm_v8x16_shuffle(v0l, v0h, 8, 24, 9, 25, 10, 26, 11, 27, 12, 28, 13, 29, 14, 30, 15, 31);
// add high bit
const v128_t v0lf = wasm_v128_or(v0lz, qhl);
const v128_t v0hf = wasm_v128_or(v0hz, qhh);
// load y
const v128_t v1l = wasm_v128_load(y0->qs);
const v128_t v1h = wasm_v128_load(y0->qs + 16);
// int8x16 -> int16x8
const v128_t v0lfl = wasm_i16x8_extend_low_i8x16 (v0lf);
const v128_t v0lfh = wasm_i16x8_extend_high_i8x16(v0lf);
const v128_t v0hfl = wasm_i16x8_extend_low_i8x16 (v0hf);
const v128_t v0hfh = wasm_i16x8_extend_high_i8x16(v0hf);
const v128_t v1ll = wasm_i16x8_extend_low_i8x16 (v1l);
const v128_t v1lh = wasm_i16x8_extend_high_i8x16(v1l);
const v128_t v1hl = wasm_i16x8_extend_low_i8x16 (v1h);
const v128_t v1hh = wasm_i16x8_extend_high_i8x16(v1h);
const float x0d = GGML_FP16_TO_FP32(x0->d);
// dot product
sumv = wasm_f32x4_add(sumv, wasm_f32x4_mul(wasm_f32x4_convert_i32x4(
wasm_i32x4_add(
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0lfl, v1ll),
wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
wasm_i32x4_dot_i16x8(v0hfh, v1hh)))), wasm_f32x4_splat(x0d*y0->d)));
}
*s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
wasm_f32x4_extract_lane(sumv, 2) + wasm_f32x4_extract_lane(sumv, 3) + summs;
#elif defined(__AVX2__)
// Initialize accumulator with zeros
__m256 acc = _mm256_setzero_ps();
@ -3827,6 +3964,7 @@ static const char * GGML_OP_LABEL[GGML_OP_COUNT] = {
"DIAG_MASK_INF",
"SOFT_MAX",
"ROPE",
"ALIBI",
"CONV_1D_1S",
"CONV_1D_2S",
@ -3875,6 +4013,7 @@ static const char * GGML_OP_SYMBOL[GGML_OP_COUNT] = {
"diag_mask_inf(x)",
"soft_max(x)",
"rope(x)",
"alibi(x)",
"conv_1d_1s(x)",
"conv_1d_2s(x)",
@ -4055,6 +4194,27 @@ bool ggml_is_quantized(enum ggml_type type) {
return GGML_IS_QUANTIZED[type];
}
enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype) {
enum ggml_type wtype = GGML_TYPE_COUNT;
switch (ftype) {
case GGML_FTYPE_ALL_F32: wtype = GGML_TYPE_F32; break;
case GGML_FTYPE_MOSTLY_F16: wtype = GGML_TYPE_F16; break;
case GGML_FTYPE_MOSTLY_Q4_0: wtype = GGML_TYPE_Q4_0; break;
case GGML_FTYPE_MOSTLY_Q4_1: wtype = GGML_TYPE_Q4_1; break;
case GGML_FTYPE_MOSTLY_Q4_2: wtype = GGML_TYPE_Q4_2; break;
case GGML_FTYPE_MOSTLY_Q5_0: wtype = GGML_TYPE_Q5_0; break;
case GGML_FTYPE_MOSTLY_Q5_1: wtype = GGML_TYPE_Q5_1; break;
case GGML_FTYPE_MOSTLY_Q8_0: wtype = GGML_TYPE_Q8_0; break;
case GGML_FTYPE_UNKNOWN: wtype = GGML_TYPE_COUNT; break;
case GGML_FTYPE_MOSTLY_Q4_1_SOME_F16: wtype = GGML_TYPE_COUNT; break;
}
GGML_ASSERT(wtype != GGML_TYPE_COUNT);
return wtype;
}
static inline bool ggml_is_transposed(const struct ggml_tensor * tensor) {
return tensor->nb[0] > tensor->nb[1];
}

16
ggml.h
View File

@ -232,6 +232,20 @@ extern "C" {
GGML_TYPE_COUNT,
};
// model file types
enum ggml_ftype {
GGML_FTYPE_UNKNOWN = -1,
GGML_FTYPE_ALL_F32 = 0,
GGML_FTYPE_MOSTLY_F16 = 1, // except 1d tensors
GGML_FTYPE_MOSTLY_Q4_0 = 2, // except 1d tensors
GGML_FTYPE_MOSTLY_Q4_1 = 3, // except 1d tensors
GGML_FTYPE_MOSTLY_Q4_1_SOME_F16 = 4, // tok_embeddings.weight and output.weight are F16
GGML_FTYPE_MOSTLY_Q4_2 = 5, // except 1d tensors
GGML_FTYPE_MOSTLY_Q8_0 = 7, // except 1d tensors
GGML_FTYPE_MOSTLY_Q5_0 = 8, // except 1d tensors
GGML_FTYPE_MOSTLY_Q5_1 = 9, // except 1d tensors
};
// available tensor operations:
enum ggml_op {
GGML_OP_NONE = 0,
@ -385,6 +399,8 @@ extern "C" {
GGML_API bool ggml_is_quantized(enum ggml_type type);
GGML_API enum ggml_type ggml_ftype_to_ggml_type(enum ggml_ftype ftype);
// main
GGML_API struct ggml_context * ggml_init(struct ggml_init_params params);

432
whisper.cpp
View File

@ -1,4 +1,3 @@
#define WHISPER_BUILD
#include "whisper.h"
#if WHISPER_USE_COREML
#include "coreml/whisper-encoder.h"
@ -255,12 +254,70 @@ 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_F32,
{
{ MODEL_TINY, 74ull*MB },
{ MODEL_BASE, 142ull*MB },
{ MODEL_SMALL, 466ull*MB },
{ MODEL_MEDIUM, 1464ull*MB },
{ MODEL_LARGE, 2952ull*MB },
},
},
{ 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 },
},
},
{ GGML_TYPE_Q4_2,
{
{ MODEL_TINY, 26ull*MB },
{ MODEL_BASE, 50ull*MB },
{ MODEL_SMALL, 154ull*MB },
{ MODEL_MEDIUM, 470ull*MB },
{ MODEL_LARGE, 940ull*MB },
},
},
{ GGML_TYPE_Q5_0, // TODO: fix
{
{ MODEL_TINY, 31ull*MB },
{ MODEL_BASE, 57ull*MB },
{ MODEL_SMALL, 181ull*MB },
{ MODEL_MEDIUM, 559ull*MB },
{ MODEL_LARGE, 1122ull*MB },
},
},
{ GGML_TYPE_Q5_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 = {
@ -370,7 +427,7 @@ struct whisper_hparams {
int32_t n_text_head = 6;
int32_t n_text_layer = 4;
int32_t n_mels = 80;
int32_t f16 = 1;
int32_t ftype = 1;
};
// audio encoding layer
@ -637,10 +694,11 @@ struct whisper_state {
};
struct whisper_context {
int64_t t_load_us = 0;
int64_t t_load_us = 0;
int64_t t_start_us = 0;
ggml_type wtype = ggml_type::GGML_TYPE_F16; // weight type (FP32 or FP16)
ggml_type wtype = ggml_type::GGML_TYPE_F16; // weight type (FP32 / FP16 / QX)
ggml_type itype = ggml_type::GGML_TYPE_F16; // intermediate type (FP32 or FP16)
whisper_model model;
whisper_vocab vocab;
@ -697,7 +755,7 @@ 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 = {
/*.mem_size =*/ cache.buf.size(),
@ -770,7 +828,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
read_safe(loader, hparams.n_text_head);
read_safe(loader, hparams.n_text_layer);
read_safe(loader, hparams.n_mels);
read_safe(loader, hparams.f16);
read_safe(loader, hparams.ftype);
assert(hparams.n_text_state == hparams.n_audio_state);
@ -794,11 +852,15 @@ 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_ftype_to_ggml_type((ggml_ftype) (model.hparams.ftype));
if (wctx.wtype == GGML_TYPE_COUNT) {
fprintf(stderr, "%s: invalid model (bad ftype value %d)\n", __func__, model.hparams.ftype);
return false;
}
const size_t scale = model.hparams.f16 ? 1 : 2;
const size_t scale = model.hparams.ftype ? 1 : 2;
fprintf(stderr, "%s: n_vocab = %d\n", __func__, hparams.n_vocab);
fprintf(stderr, "%s: n_audio_ctx = %d\n", __func__, hparams.n_audio_ctx);
@ -810,18 +872,18 @@ 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 = %d\n", __func__, model.hparams.ftype);
fprintf(stderr, "%s: type = %d\n", __func__, model.type);
// print memory requirements
{
// this is the total memory required to run the inference
const size_t mem_required =
MEM_REQ_SCRATCH0.at (model.type) +
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) +
MEM_REQ_SCRATCH0.at(model.type) +
MEM_REQ_SCRATCH1.at(model.type) +
MEM_REQ_SCRATCH2.at(model.type) +
MEM_REQ_SCRATCH3.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));
@ -837,7 +899,7 @@ 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));
// we skip initialization of the state until it is needed
// because it might be that state will always be provided externally.
@ -928,6 +990,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
size_t ctx_size = 0;
const ggml_type wtype = wctx.wtype;
const ggml_type vtype = wctx.wtype == GGML_TYPE_F32 ? GGML_TYPE_F32 : GGML_TYPE_F16; // conv type
{
const auto & hparams = model.hparams;
@ -946,92 +1009,92 @@ 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(vtype); // 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(vtype); // 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
@ -1077,175 +1140,175 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
// encoder
{
model.e_pe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_audio_state, n_audio_ctx);
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, vtype, 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, vtype, 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);
model.e_ln_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
model.e_ln_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
model.e_ln_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
// map by name
model.tensors["encoder.positional_embedding"] = model.e_pe;
model.tensors["encoder.conv1.weight"] = model.e_conv_1_w;
model.tensors["encoder.conv1.bias"] = model.e_conv_1_b;
model.tensors["encoder.conv1.weight"] = model.e_conv_1_w;
model.tensors["encoder.conv1.bias"] = model.e_conv_1_b;
model.tensors["encoder.conv2.weight"] = model.e_conv_2_w;
model.tensors["encoder.conv2.bias"] = model.e_conv_2_b;
model.tensors["encoder.conv2.weight"] = model.e_conv_2_w;
model.tensors["encoder.conv2.bias"] = model.e_conv_2_b;
model.tensors["encoder.ln_post.weight"] = model.e_ln_w;
model.tensors["encoder.ln_post.bias"] = model.e_ln_b;
model.tensors["encoder.ln_post.weight"] = model.e_ln_w;
model.tensors["encoder.ln_post.bias"] = model.e_ln_b;
for (int i = 0; i < n_audio_layer; ++i) {
auto & layer = model.layers_encoder[i];
layer.mlp_ln_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.mlp_ln_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.mlp_ln_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.mlp_ln_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.mlp_0_w = ggml_new_tensor_2d(ctx, wtype, n_audio_state, 4*n_audio_state);
layer.mlp_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_audio_state);
layer.mlp_0_w = ggml_new_tensor_2d(ctx, wtype, n_audio_state, 4*n_audio_state);
layer.mlp_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_audio_state);
layer.mlp_1_w = ggml_new_tensor_2d(ctx, wtype, 4*n_audio_state, n_audio_state);
layer.mlp_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.mlp_1_w = ggml_new_tensor_2d(ctx, wtype, 4*n_audio_state, n_audio_state);
layer.mlp_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.attn_ln_0_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.attn_ln_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.attn_ln_0_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.attn_ln_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.attn_q_w = ggml_new_tensor_2d(ctx, wtype, n_audio_state, n_audio_state);
layer.attn_q_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.attn_q_w = ggml_new_tensor_2d(ctx, wtype, n_audio_state, n_audio_state);
layer.attn_q_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.attn_k_w = ggml_new_tensor_2d(ctx, wtype, n_audio_state, n_audio_state);
layer.attn_k_w = ggml_new_tensor_2d(ctx, wtype, n_audio_state, n_audio_state);
layer.attn_v_w = ggml_new_tensor_2d(ctx, wtype, n_audio_state, n_audio_state);
layer.attn_v_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.attn_v_w = ggml_new_tensor_2d(ctx, wtype, n_audio_state, n_audio_state);
layer.attn_v_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.attn_ln_1_w = ggml_new_tensor_2d(ctx, wtype, n_audio_state, n_audio_state);
layer.attn_ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
layer.attn_ln_1_w = ggml_new_tensor_2d(ctx, wtype, n_audio_state, n_audio_state);
layer.attn_ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_audio_state);
// map by name
model.tensors["encoder.blocks." + std::to_string(i) + ".mlp_ln.weight"] = layer.mlp_ln_w;
model.tensors["encoder.blocks." + std::to_string(i) + ".mlp_ln.bias"] = layer.mlp_ln_b;
model.tensors["encoder.blocks." + std::to_string(i) + ".mlp_ln.weight"] = layer.mlp_ln_w;
model.tensors["encoder.blocks." + std::to_string(i) + ".mlp_ln.bias"] = layer.mlp_ln_b;
model.tensors["encoder.blocks." + std::to_string(i) + ".mlp.0.weight"] = layer.mlp_0_w;
model.tensors["encoder.blocks." + std::to_string(i) + ".mlp.0.bias"] = layer.mlp_0_b;
model.tensors["encoder.blocks." + std::to_string(i) + ".mlp.0.weight"] = layer.mlp_0_w;
model.tensors["encoder.blocks." + std::to_string(i) + ".mlp.0.bias"] = layer.mlp_0_b;
model.tensors["encoder.blocks." + std::to_string(i) + ".mlp.2.weight"] = layer.mlp_1_w;
model.tensors["encoder.blocks." + std::to_string(i) + ".mlp.2.bias"] = layer.mlp_1_b;
model.tensors["encoder.blocks." + std::to_string(i) + ".mlp.2.weight"] = layer.mlp_1_w;
model.tensors["encoder.blocks." + std::to_string(i) + ".mlp.2.bias"] = layer.mlp_1_b;
model.tensors["encoder.blocks." + std::to_string(i) + ".attn_ln.weight"] = layer.attn_ln_0_w;
model.tensors["encoder.blocks." + std::to_string(i) + ".attn_ln.bias"] = layer.attn_ln_0_b;
model.tensors["encoder.blocks." + std::to_string(i) + ".attn_ln.weight"] = layer.attn_ln_0_w;
model.tensors["encoder.blocks." + std::to_string(i) + ".attn_ln.bias"] = layer.attn_ln_0_b;
model.tensors["encoder.blocks." + std::to_string(i) + ".attn.query.weight"] = layer.attn_q_w;
model.tensors["encoder.blocks." + std::to_string(i) + ".attn.query.bias"] = layer.attn_q_b;
model.tensors["encoder.blocks." + std::to_string(i) + ".attn.key.weight"] = layer.attn_k_w;
model.tensors["encoder.blocks." + std::to_string(i) + ".attn.key.weight"] = layer.attn_k_w;
model.tensors["encoder.blocks." + std::to_string(i) + ".attn.value.weight"] = layer.attn_v_w;
model.tensors["encoder.blocks." + std::to_string(i) + ".attn.value.bias"] = layer.attn_v_b;
model.tensors["encoder.blocks." + std::to_string(i) + ".attn.out.weight"] = layer.attn_ln_1_w;
model.tensors["encoder.blocks." + std::to_string(i) + ".attn.out.bias"] = layer.attn_ln_1_b;
model.tensors["encoder.blocks." + std::to_string(i) + ".attn.out.weight"] = layer.attn_ln_1_w;
model.tensors["encoder.blocks." + std::to_string(i) + ".attn.out.bias"] = layer.attn_ln_1_b;
}
}
// decoder
{
model.d_pe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_text_state, n_text_ctx);
model.d_pe = ggml_new_tensor_2d(ctx, GGML_TYPE_F32, n_text_state, n_text_ctx);
model.d_te = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_vocab);
model.d_te = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_vocab);
model.d_ln_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
model.d_ln_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
// map by name
model.tensors["decoder.positional_embedding"] = model.d_pe;
model.tensors["decoder.positional_embedding"] = model.d_pe;
model.tensors["decoder.token_embedding.weight"] = model.d_te;
model.tensors["decoder.ln.weight"] = model.d_ln_w;
model.tensors["decoder.ln.bias"] = model.d_ln_b;
model.tensors["decoder.ln.weight"] = model.d_ln_w;
model.tensors["decoder.ln.bias"] = model.d_ln_b;
for (int i = 0; i < n_text_layer; ++i) {
auto & layer = model.layers_decoder[i];
layer.mlp_ln_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.mlp_ln_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.mlp_ln_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.mlp_ln_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.mlp_0_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, 4*n_text_state);
layer.mlp_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_text_state);
layer.mlp_0_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, 4*n_text_state);
layer.mlp_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, 4*n_text_state);
layer.mlp_1_w = ggml_new_tensor_2d(ctx, wtype, 4*n_text_state, n_text_state);
layer.mlp_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.mlp_1_w = ggml_new_tensor_2d(ctx, wtype, 4*n_text_state, n_text_state);
layer.mlp_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.attn_ln_0_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.attn_ln_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.attn_ln_0_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.attn_ln_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.attn_q_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.attn_q_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.attn_q_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.attn_q_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.attn_k_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.attn_k_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.attn_v_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.attn_v_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.attn_v_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.attn_v_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.attn_ln_1_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.attn_ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.attn_ln_1_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.attn_ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.cross_attn_ln_0_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.cross_attn_ln_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.cross_attn_ln_0_w = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.cross_attn_ln_0_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.cross_attn_q_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.cross_attn_q_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.cross_attn_q_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.cross_attn_q_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.cross_attn_k_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.cross_attn_k_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.cross_attn_v_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.cross_attn_v_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.cross_attn_v_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.cross_attn_v_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.cross_attn_ln_1_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.cross_attn_ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
layer.cross_attn_ln_1_w = ggml_new_tensor_2d(ctx, wtype, n_text_state, n_text_state);
layer.cross_attn_ln_1_b = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_text_state);
// map by name
model.tensors["decoder.blocks." + std::to_string(i) + ".mlp_ln.weight"] = layer.mlp_ln_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".mlp_ln.bias"] = layer.mlp_ln_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".mlp_ln.weight"] = layer.mlp_ln_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".mlp_ln.bias"] = layer.mlp_ln_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".mlp.0.weight"] = layer.mlp_0_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".mlp.0.bias"] = layer.mlp_0_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".mlp.0.weight"] = layer.mlp_0_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".mlp.0.bias"] = layer.mlp_0_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".mlp.2.weight"] = layer.mlp_1_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".mlp.2.bias"] = layer.mlp_1_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".mlp.2.weight"] = layer.mlp_1_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".mlp.2.bias"] = layer.mlp_1_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn_ln.weight"] = layer.attn_ln_0_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn_ln.bias"] = layer.attn_ln_0_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn_ln.weight"] = layer.attn_ln_0_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn_ln.bias"] = layer.attn_ln_0_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn.query.weight"] = layer.attn_q_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn.query.bias"] = layer.attn_q_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn.query.weight"] = layer.attn_q_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn.query.bias"] = layer.attn_q_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn.key.weight"] = layer.attn_k_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn.key.weight"] = layer.attn_k_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn.value.weight"] = layer.attn_v_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn.value.bias"] = layer.attn_v_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn.value.weight"] = layer.attn_v_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn.value.bias"] = layer.attn_v_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn.out.weight"] = layer.attn_ln_1_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn.out.bias"] = layer.attn_ln_1_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn.out.weight"] = layer.attn_ln_1_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".attn.out.bias"] = layer.attn_ln_1_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn_ln.weight"] = layer.cross_attn_ln_0_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn_ln.bias"] = layer.cross_attn_ln_0_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn_ln.weight"] = layer.cross_attn_ln_0_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn_ln.bias"] = layer.cross_attn_ln_0_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.query.weight"] = layer.cross_attn_q_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.query.bias"] = layer.cross_attn_q_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.key.weight"] = layer.cross_attn_k_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.key.weight"] = layer.cross_attn_k_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.value.weight"] = layer.cross_attn_v_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.value.bias"] = layer.cross_attn_v_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.out.weight"] = layer.cross_attn_ln_1_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.out.bias"] = layer.cross_attn_ln_1_b;
model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.out.weight"] = layer.cross_attn_ln_1_w;
model.tensors["decoder.blocks." + std::to_string(i) + ".cross_attn.out.bias"] = layer.cross_attn_ln_1_b;
}
}
}
@ -1259,11 +1322,11 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
while (true) {
int32_t n_dims;
int32_t length;
int32_t ftype;
int32_t ttype;
read_safe(loader, n_dims);
read_safe(loader, length);
read_safe(loader, ftype);
read_safe(loader, ttype);
if (loader->eof(loader->context)) {
break;
@ -1298,9 +1361,9 @@ 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);
const size_t bpe = ggml_type_size(ggml_type(ttype));
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;
@ -1309,7 +1372,7 @@ static bool whisper_model_load(struct whisper_model_loader * loader, whisper_con
loader->read(loader->context, tensor->data, ggml_nbytes(tensor));
BYTESWAP_TENSOR(tensor);
//printf("%48s - [%5d, %5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ne[2], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
//printf("%48s - [%5d, %5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ne[2], ggml_type_name((ggml_type) ttype), ggml_nbytes(tensor)/1024.0/1024.0);
total_size += ggml_nbytes(tensor);
model.n_loaded++;
}
@ -1508,14 +1571,14 @@ static bool whisper_encode_internal(
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, wctx.itype, 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, wctx.itype, n_state/n_head, n_head, n_ctx)),
0, 2, 1, 3);
struct ggml_tensor * V =
@ -1525,7 +1588,7 @@ static bool whisper_encode_internal(
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, wctx.itype, n_ctx, n_state/n_head, n_head));
struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, false);
#else
@ -1540,7 +1603,7 @@ static bool whisper_encode_internal(
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, wctx.itype, n_state/n_head, n_head, n_ctx)),
0, 2, 1, 3);
// K * Q
@ -1561,7 +1624,7 @@ static bool whisper_encode_internal(
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, wctx.itype, n_ctx, n_state/n_head, n_head)
);
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V, KQ_soft_max);
@ -1619,7 +1682,7 @@ static bool whisper_encode_internal(
wstate.use_buf(ctx0, 0);
cur = ggml_flash_ff(ctx0,
ggml_cpy(ctx0, cur, ggml_new_tensor_2d(ctx0, wstate.wtype, n_state, n_ctx)),
ggml_cpy(ctx0, cur, ggml_new_tensor_2d(ctx0, wstate.itype, n_state, n_ctx)),
layer.mlp_0_w, layer.mlp_0_b, layer.mlp_1_w, layer.mlp_1_b);
#else
wstate.use_buf(ctx0, 0);
@ -2537,9 +2600,9 @@ static std::string whisper_get_coreml_path_encoder(std::string path_bin) {
struct whisper_state * whisper_init_state(whisper_context * ctx) {
whisper_state * state = new whisper_state;
const size_t scale = ctx->model.hparams.f16 ? 1 : 2;
const size_t scale = ctx->model.hparams.ftype ? 1 : 2;
if (!kv_cache_init(ctx->model.hparams, scale * MEM_REQ_KV_SELF.at(ctx->model.type), state->decoders[0].kv_self, ctx->wtype, ctx->model.hparams.n_text_ctx)) {
if (!kv_cache_init(ctx->model.hparams, scale * MEM_REQ_KV_SELF.at(ctx->model.type), state->decoders[0].kv_self, ctx->itype, ctx->model.hparams.n_text_ctx)) {
fprintf(stderr, "%s: kv_cache_init() failed for self-attention cache\n", __func__);
delete state;
return nullptr;
@ -2550,7 +2613,7 @@ struct whisper_state * whisper_init_state(whisper_context * ctx) {
fprintf(stderr, "%s: kv self size = %7.2f MB\n", __func__, memory_size / 1024.0 / 1024.0);
}
if (!kv_cache_init(ctx->model.hparams, scale * MEM_REQ_KV_CROSS.at(ctx->model.type), state->kv_cross, ctx->wtype, ctx->model.hparams.n_audio_ctx)) {
if (!kv_cache_init(ctx->model.hparams, scale * MEM_REQ_KV_CROSS.at(ctx->model.type), state->kv_cross, ctx->itype, ctx->model.hparams.n_audio_ctx)) {
fprintf(stderr, "%s: kv_cache_init() failed for cross-attention cache\n", __func__);
delete state;
return nullptr;
@ -3049,8 +3112,8 @@ int whisper_model_n_mels(struct whisper_context * ctx) {
return ctx->model.hparams.n_mels;
}
int whisper_model_f16(struct whisper_context * ctx) {
return ctx->model.hparams.f16;
int whisper_model_ftype(struct whisper_context * ctx) {
return ctx->model.hparams.ftype;
}
int whisper_model_type(struct whisper_context * ctx) {
@ -4825,23 +4888,32 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
// when F16 is used, there is an extra work buffer of size N*N*sizeof(float)
std::vector<char> buf(4llu*N_max*N_max*sizeof(float) + 4*256);
// put a bunch of random data in the buffer
for (size_t i = 0; i < buf.size(); i++) buf[i] = i;
for (int j = 0; j < (int) sizes.size(); j++) {
int n_q4_0 = 0;
int n_q4_1 = 0;
int n_fp16 = 0;
int n_fp32 = 0;
// GFLOPS/s
double s_q4_0 = 0.0;
double s_q4_1 = 0.0;
double s_fp16 = 0.0;
double s_fp32 = 0.0;
const size_t N = sizes[j];
for (int k = 0; k < 2; ++k) {
const ggml_type wtype = k == 0 ? GGML_TYPE_F16 : GGML_TYPE_F32;
for (int k = 0; k < 4; ++k) {
const ggml_type wtype =
k == 0 ? GGML_TYPE_Q4_0 :
k == 1 ? GGML_TYPE_Q4_1 :
k == 2 ? GGML_TYPE_F16 :
GGML_TYPE_F32;
double & s = k == 0 ? s_fp16 : s_fp32;
int & n = k == 0 ? n_fp16 : n_fp32;
double & s = k == 0 ? s_q4_0 : k == 1 ? s_q4_1 : k == 2 ? s_fp16 : s_fp32;
int & n = k == 0 ? n_q4_0 : k == 1 ? n_q4_1 : k == 2 ? n_fp16 : n_fp32;
struct ggml_init_params gparams = {
/*.mem_size =*/ buf.size(),
@ -4885,8 +4957,8 @@ WHISPER_API const char * whisper_bench_ggml_mul_mat_str(int n_threads) {
s = ((2.0*N*N*N*n)/tsum)*1e-9;
}
snprintf(strbuf, sizeof(strbuf), "ggml_mul_mat: %5zu x %5zu: F16 %8.1f GFLOPS (%3d runs) / F32 %8.1f GFLOPS (%3d runs)\n",
N, N, s_fp16, n_fp16, s_fp32, n_fp32);
snprintf(strbuf, sizeof(strbuf), "ggml_mul_mat: %4zu x %4zu: Q4_0 %7.1f GFLOPS (%3d runs) / Q4_1 %7.1f GFLOPS (%3d runs) / F16 %7.1f GFLOPS (%3d runs) / F32 %7.1f GFLOPS (%3d runs)\n",
N, N, s_q4_0, n_q4_0, s_q4_1, n_q4_1, s_fp16, n_fp16, s_fp32, n_fp32);
s += strbuf;
}

2
whisper.h
View File

@ -258,7 +258,7 @@ extern "C" {
WHISPER_API int whisper_model_n_text_head (struct whisper_context * ctx);
WHISPER_API int whisper_model_n_text_layer (struct whisper_context * ctx);
WHISPER_API int whisper_model_n_mels (struct whisper_context * ctx);
WHISPER_API int whisper_model_f16 (struct whisper_context * ctx);
WHISPER_API int whisper_model_ftype (struct whisper_context * ctx);
WHISPER_API int whisper_model_type (struct whisper_context * ctx);
// Token logits obtained from the last call to whisper_decode()