talk : initial real-time transcription in the browser

pull/155/head
Georgi Gerganov 3 years ago
parent 2e311a2917
commit 6f110d5425
No known key found for this signature in database
GPG Key ID: 449E073F9DC10735

@ -9,12 +9,13 @@ target_link_libraries(${TARGET} PRIVATE
) )
unset(EXTRA_FLAGS) unset(EXTRA_FLAGS)
if (WHISPER_WASM_SINGLE_FILE) if (WHISPER_WASM_SINGLE_FILE)
set(EXTRA_FLAGS "-s SINGLE_FILE=1") set(EXTRA_FLAGS "-s SINGLE_FILE=1")
message(STATUS "Embedding WASM inside whisper.js") message(STATUS "Embedding WASM inside whisper.js")
add_custom_command( add_custom_command(
TARGET libwhisper POST_BUILD TARGET ${TARGET} POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy COMMAND ${CMAKE_COMMAND} -E copy
${CMAKE_BINARY_DIR}/bin/libwhisper.js ${CMAKE_BINARY_DIR}/bin/libwhisper.js
${CMAKE_CURRENT_SOURCE_DIR}/whisper.js ${CMAKE_CURRENT_SOURCE_DIR}/whisper.js

File diff suppressed because one or more lines are too long

@ -20,6 +20,7 @@ include_directories(${CMAKE_CURRENT_SOURCE_DIR})
if (EMSCRIPTEN) if (EMSCRIPTEN)
add_subdirectory(whisper.wasm) add_subdirectory(whisper.wasm)
add_subdirectory(talk)
else() else()
add_subdirectory(main) add_subdirectory(main)
add_subdirectory(stream) add_subdirectory(stream)

@ -0,0 +1,46 @@
#
# libtalk
#
set(TARGET libtalk)
add_executable(${TARGET}
emscripten.cpp
)
target_link_libraries(${TARGET} PRIVATE
whisper
)
unset(EXTRA_FLAGS)
if (WHISPER_WASM_SINGLE_FILE)
set(EXTRA_FLAGS "-s SINGLE_FILE=1")
message(STATUS "Embedding WASM inside talk.js")
add_custom_command(
TARGET ${TARGET} POST_BUILD
COMMAND ${CMAKE_COMMAND} -E copy
${CMAKE_BINARY_DIR}/bin/libtalk.js
${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/talk/talk.js
)
endif()
set_target_properties(${TARGET} PROPERTIES LINK_FLAGS " \
--bind \
-s USE_PTHREADS=1 \
-s PTHREAD_POOL_SIZE=8 \
-s INITIAL_MEMORY=1400MB \
-s TOTAL_MEMORY=1400MB \
-s FORCE_FILESYSTEM=1 \
-s EXPORTED_RUNTIME_METHODS=\"['print', 'printErr', 'ccall', 'cwrap']\" \
${EXTRA_FLAGS} \
")
#
# talk
#
set(TARGET talk)
configure_file(${CMAKE_CURRENT_SOURCE_DIR}/index-tmpl.html ${CMAKE_RUNTIME_OUTPUT_DIRECTORY}/${TARGET}/index.html @ONLY)

@ -0,0 +1,7 @@
# talk
WIP IN PROGRESS
ref: https://github.com/ggerganov/whisper.cpp/issues/154
demo: https://talk.ggerganov.com

File diff suppressed because it is too large Load Diff

@ -0,0 +1,793 @@
#include "ggml/ggml.h"
#include "utils.h"
#include <cassert>
#include <cmath>
#include <cstdio>
#include <cstring>
#include <fstream>
#include <map>
#include <string>
#include <vector>
// default hparams (GPT-2 117M)
struct gpt2_hparams {
int32_t n_vocab = 50257;
int32_t n_ctx = 1024;
int32_t n_embd = 768;
int32_t n_head = 12;
int32_t n_layer = 12;
int32_t f16 = 1;
};
struct gpt2_layer {
// normalization
struct ggml_tensor * ln_1_g;
struct ggml_tensor * ln_1_b;
struct ggml_tensor * ln_2_g;
struct ggml_tensor * ln_2_b;
// attention
struct ggml_tensor * c_attn_attn_w;
struct ggml_tensor * c_attn_attn_b;
struct ggml_tensor * c_attn_proj_w;
struct ggml_tensor * c_attn_proj_b;
// mlp
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_b;
};
struct gpt2_model {
gpt2_hparams hparams;
// normalization
struct ggml_tensor * ln_f_g;
struct ggml_tensor * ln_f_b;
struct ggml_tensor * wte; // position embedding
struct ggml_tensor * wpe; // token embedding
std::vector<gpt2_layer> layers;
// key + value memory
struct ggml_tensor * memory_k;
struct ggml_tensor * memory_v;
//
struct ggml_context * ctx;
std::map<std::string, struct ggml_tensor *> tensors;
};
// load the model's weights from a file
bool gpt2_model_load(const std::string & fname, gpt2_model & model, gpt_vocab & vocab) {
printf("%s: loading model from '%s'\n", __func__, fname.c_str());
auto fin = std::ifstream(fname, std::ios::binary);
if (!fin) {
fprintf(stderr, "%s: failed to open '%s'\n", __func__, fname.c_str());
return false;
}
// verify magic
{
uint32_t magic;
fin.read((char *) &magic, sizeof(magic));
if (magic != 0x67676d6c) {
fprintf(stderr, "%s: invalid model file '%s' (bad magic)\n", __func__, fname.c_str());
return false;
}
}
// load hparams
{
auto & hparams = model.hparams;
fin.read((char *) &hparams.n_vocab, sizeof(hparams.n_vocab));
fin.read((char *) &hparams.n_ctx, sizeof(hparams.n_ctx));
fin.read((char *) &hparams.n_embd, sizeof(hparams.n_embd));
fin.read((char *) &hparams.n_head, sizeof(hparams.n_head));
fin.read((char *) &hparams.n_layer, sizeof(hparams.n_layer));
fin.read((char *) &hparams.f16, sizeof(hparams.f16));
printf("%s: n_vocab = %d\n", __func__, hparams.n_vocab);
printf("%s: n_ctx = %d\n", __func__, hparams.n_ctx);
printf("%s: n_embd = %d\n", __func__, hparams.n_embd);
printf("%s: n_head = %d\n", __func__, hparams.n_head);
printf("%s: n_layer = %d\n", __func__, hparams.n_layer);
printf("%s: f16 = %d\n", __func__, hparams.f16);
}
// load vocab
{
int32_t n_vocab = 0;
fin.read((char *) &n_vocab, sizeof(n_vocab));
if (n_vocab != model.hparams.n_vocab) {
fprintf(stderr, "%s: invalid model file '%s' (bad vocab size %d != %d)\n",
__func__, fname.c_str(), n_vocab, model.hparams.n_vocab);
return false;
}
std::string word;
for (int i = 0; i < n_vocab; i++) {
uint32_t len;
fin.read((char *) &len, sizeof(len));
word.resize(len);
fin.read((char *) word.data(), len);
vocab.token_to_id[word] = i;
vocab.id_to_token[i] = word;
}
}
// for the big tensors, we have the option to store the data in 16-bit floats
// 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;
auto & ctx = model.ctx;
size_t ctx_size = 0;
{
const auto & hparams = model.hparams;
const int n_embd = hparams.n_embd;
const int n_layer = hparams.n_layer;
const int n_ctx = hparams.n_ctx;
const int n_vocab = hparams.n_vocab;
ctx_size += n_embd*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_vocab*n_embd*ggml_type_size(wtype); // wte
ctx_size += n_ctx*n_embd*ggml_type_size(GGML_TYPE_F32); // wpe
ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_g
ctx_size += n_layer*(n_embd*ggml_type_size(GGML_TYPE_F32)); // ln_1_b
ctx_size += n_layer*(n_embd*ggml_type_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*(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*(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*(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_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_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 += (6 + 12*n_layer)*256; // object overhead
printf("%s: ggml ctx size = %6.2f MB\n", __func__, ctx_size/(1024.0*1024.0));
}
// create the ggml context
{
struct ggml_init_params params = {
.mem_size = ctx_size,
.mem_buffer = NULL,
};
model.ctx = ggml_init(params);
if (!model.ctx) {
fprintf(stderr, "%s: ggml_init() failed\n", __func__);
return false;
}
}
// prepare memory for the weights
{
const auto & hparams = model.hparams;
const int n_embd = hparams.n_embd;
const int n_layer = hparams.n_layer;
const int n_ctx = hparams.n_ctx;
const int n_vocab = hparams.n_vocab;
model.layers.resize(n_layer);
model.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);
// 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;
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_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_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_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);
// map by name
model.tensors["model/h" + std::to_string(i) + "/ln_1/g"] = layer.ln_1_g;
model.tensors["model/h" + std::to_string(i) + "/ln_1/b"] = layer.ln_1_b;
model.tensors["model/h" + std::to_string(i) + "/ln_2/g"] = layer.ln_2_g;
model.tensors["model/h" + std::to_string(i) + "/ln_2/b"] = layer.ln_2_b;
model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/w"] = layer.c_attn_attn_w;
model.tensors["model/h" + std::to_string(i) + "/attn/c_attn/b"] = layer.c_attn_attn_b;
model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/w"] = layer.c_attn_proj_w;
model.tensors["model/h" + std::to_string(i) + "/attn/c_proj/b"] = layer.c_attn_proj_b;
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/b"] = layer.c_mlp_proj_b;
}
}
// key + value memory
{
const auto & hparams = model.hparams;
const int n_embd = hparams.n_embd;
const int n_layer = hparams.n_layer;
const int n_ctx = hparams.n_ctx;
const int n_mem = n_layer*n_ctx;
const int n_elements = n_embd*n_mem;
model.memory_k = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
model.memory_v = ggml_new_tensor_1d(ctx, GGML_TYPE_F32, n_elements);
const size_t memory_size = ggml_nbytes(model.memory_k) + ggml_nbytes(model.memory_v);
printf("%s: memory size = %8.2f MB, n_mem = %d\n", __func__, memory_size/1024.0/1024.0, n_mem);
}
// load weights
{
size_t total_size = 0;
while (true) {
int32_t n_dims;
int32_t length;
int32_t ftype;
fin.read(reinterpret_cast<char *>(&n_dims), sizeof(n_dims));
fin.read(reinterpret_cast<char *>(&length), sizeof(length));
fin.read(reinterpret_cast<char *>(&ftype), sizeof(ftype));
if (fin.eof()) {
break;
}
int32_t nelements = 1;
int32_t ne[2] = { 1, 1 };
for (int i = 0; i < n_dims; ++i) {
fin.read(reinterpret_cast<char *>(&ne[i]), sizeof(ne[i]));
nelements *= ne[i];
}
std::string name(length, 0);
fin.read(&name[0], length);
if (model.tensors.find(name.data()) == model.tensors.end()) {
fprintf(stderr, "%s: unknown tensor '%s' in model file\n", __func__, name.data());
return false;
}
auto tensor = model.tensors[name.data()];
if (ggml_nelements(tensor) != nelements) {
fprintf(stderr, "%s: tensor '%s' has wrong size in model file\n", __func__, name.data());
return false;
}
if (tensor->ne[0] != ne[0] || tensor->ne[1] != ne[1]) {
fprintf(stderr, "%s: tensor '%s' has wrong shape in model file: got [%d, %d], expected [%d, %d]\n",
__func__, name.data(), tensor->ne[0], tensor->ne[1], ne[0], ne[1]);
return false;
}
const size_t bpe = (ftype == 0) ? sizeof(float) : sizeof(ggml_fp16_t);
if (nelements*bpe != ggml_nbytes(tensor)) {
fprintf(stderr, "%s: tensor '%s' has wrong size in model file: got %zu, expected %zu\n",
__func__, name.data(), ggml_nbytes(tensor), nelements*bpe);
return false;
}
fin.read(reinterpret_cast<char *>(tensor->data), ggml_nbytes(tensor));
//printf("%24s - [%5d, %5d], type = %6s, %6.2f MB\n", name.data(), ne[0], ne[1], ftype == 0 ? "float" : "f16", ggml_nbytes(tensor)/1024.0/1024.0);
total_size += ggml_nbytes(tensor);
}
printf("%s: model size = %8.2f MB\n", __func__, total_size/1024.0/1024.0);
}
fin.close();
return true;
}
// evaluate the transformer
//
// - model: the model
// - n_threads: number of threads to use
// - n_past: the context size so far
// - embd_inp: the embeddings of the tokens in the context
// - embd_w: the predicted probabilities of the next token
//
bool gpt2_eval(
const gpt2_model & model,
const int n_threads,
const int n_past,
const std::vector<gpt_vocab::id> & embd_inp,
std::vector<float> & embd_w,
size_t & mem_per_token) {
const int N = embd_inp.size();
const auto & hparams = model.hparams;
const int n_embd = hparams.n_embd;
const int n_layer = hparams.n_layer;
const int n_ctx = hparams.n_ctx;
const int n_head = hparams.n_head;
const int n_vocab = hparams.n_vocab;
static size_t buf_size = 256u*1024*1024;
static void * buf = malloc(buf_size);
if (mem_per_token > 0 && mem_per_token*N > buf_size) {
const size_t buf_size_new = 1.1*(mem_per_token*N); // add 10% to account for ggml object overhead
//printf("\n%s: reallocating buffer from %zu to %zu bytes\n", __func__, buf_size, buf_size_new);
// reallocate
buf_size = buf_size_new;
buf = realloc(buf, buf_size);
if (buf == nullptr) {
fprintf(stderr, "%s: failed to allocate %zu bytes\n", __func__, buf_size);
return false;
}
}
struct ggml_init_params params = {
.mem_size = buf_size,
.mem_buffer = buf,
};
struct ggml_context * ctx0 = ggml_init(params);
struct ggml_cgraph gf = { .n_threads = n_threads };
struct ggml_tensor * embd = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
memcpy(embd->data, embd_inp.data(), N*ggml_element_size(embd));
struct ggml_tensor * position = ggml_new_tensor_1d(ctx0, GGML_TYPE_I32, N);
for (int i = 0; i < N; ++i) {
((int32_t *) position->data)[i] = n_past + i;
}
// wte + wpe
struct ggml_tensor * inpL =
ggml_add(ctx0,
ggml_get_rows(ctx0, model.wte, embd),
ggml_get_rows(ctx0, model.wpe, position));
for (int il = 0; il < n_layer; ++il) {
struct ggml_tensor * cur;
// norm
{
// [ 768, N]
cur = ggml_norm(ctx0, inpL);
// cur = ln_1_g*cur + ln_1_b
// [ 768, N]
cur = ggml_add(ctx0,
ggml_mul(ctx0,
ggml_repeat(ctx0, model.layers[il].ln_1_g, cur),
cur),
ggml_repeat(ctx0, model.layers[il].ln_1_b, cur));
}
// attn
// [2304, 768] - model.layers[il].c_attn_attn_w
// [2304, 1] - model.layers[il].c_attn_attn_b
// [ 768, N] - cur (in)
// [2304, N] - cur (out)
//
// cur = attn_w*cur + attn_b
// [2304, N]
{
cur = ggml_mul_mat(ctx0,
ggml_transpose(ctx0, model.layers[il].c_attn_attn_w),
cur);
cur = ggml_add(ctx0,
ggml_repeat(ctx0, model.layers[il].c_attn_attn_b, cur),
cur);
}
// self-attention
{
struct ggml_tensor * Qcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 0*sizeof(float)*n_embd);
struct ggml_tensor * Kcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 1*sizeof(float)*n_embd);
struct ggml_tensor * Vcur = ggml_view_2d(ctx0, cur, n_embd, N, cur->nb[1], 2*sizeof(float)*n_embd);
// store key and value to memory
if (N >= 1) {
struct ggml_tensor * k = ggml_view_1d(ctx0, model.memory_k, N*n_embd, (ggml_element_size(model.memory_k)*n_embd)*(il*n_ctx + n_past));
struct ggml_tensor * v = ggml_view_1d(ctx0, model.memory_v, N*n_embd, (ggml_element_size(model.memory_v)*n_embd)*(il*n_ctx + n_past));
ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Kcur, k));
ggml_build_forward_expand(&gf, ggml_cpy(ctx0, Vcur, v));
}
// Q = Qcur.contiguous().view(n_embd/n_head, n_head, N).permute(0, 2, 1, 3)
// [64, N, 12]
struct ggml_tensor * Q =
ggml_permute(ctx0,
ggml_cpy(ctx0,
Qcur,
ggml_new_tensor_3d(ctx0, GGML_TYPE_F32, n_embd/n_head, n_head, N)),
0, 2, 1, 3);
// K = Kmem.view(n_embd/n_head, n_head, n_past + N).permute(0, 2, 1, 3)
// [64, n_past + N, 12]
struct ggml_tensor * K =
ggml_permute(ctx0,
ggml_reshape_3d(ctx0,
ggml_view_1d(ctx0, model.memory_k, (n_past + N)*n_embd, il*n_ctx*ggml_element_size(model.memory_k)*n_embd),
n_embd/n_head, n_head, n_past + N),
0, 2, 1, 3);
// GG: flash attention
//struct ggml_tensor * V =
// 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, GGML_TYPE_F32, n_past + N, n_embd/n_head, n_head));
//struct ggml_tensor * KQV = ggml_flash_attn(ctx0, Q, K, V, true);
// K * Q
// [n_past + N, N, 12]
struct ggml_tensor * KQ = ggml_mul_mat(ctx0, K, Q);
// KQ_scaled = KQ / sqrt(n_embd/n_head)
// [n_past + N, N, 12]
struct ggml_tensor * KQ_scaled =
ggml_scale(ctx0,
KQ,
ggml_new_f32(ctx0, 1.0f/sqrt(float(n_embd)/n_head))
);
// KQ_masked = mask_past(KQ_scaled)
// [n_past + N, N, 12]
struct ggml_tensor * KQ_masked = ggml_diag_mask_inf(ctx0, KQ_scaled, n_past);
// KQ = soft_max(KQ_masked)
// [n_past + N, N, 12]
struct ggml_tensor * KQ_soft_max = ggml_soft_max(ctx0, KQ_masked);
// V_trans = Vmem.view(n_embd/n_head, n_head, n_past + N).permute(1, 2, 0, 3).contiguous()
// [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);
// KQV = transpose(V) * KQ_soft_max
// [64, N, 12]
struct ggml_tensor * KQV = ggml_mul_mat(ctx0, V_trans, KQ_soft_max);
// KQV_merged = KQV.permute(0, 2, 1, 3)
// [64, 12, N]
struct ggml_tensor * KQV_merged = ggml_permute(ctx0, KQV, 0, 2, 1, 3);
// cur = KQV_merged.contiguous().view(n_embd, N)
// [768, N]
cur = ggml_cpy(ctx0,
KQV_merged,
ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_embd, N));
}
// projection
// [ 768, 768] - model.layers[il].c_attn_proj_w
// [ 768, 1] - model.layers[il].c_attn_proj_b
// [ 768, N] - cur (in)
// [ 768, N] - cur (out)
//
// cur = proj_w*cur + proj_b
// [768, N]
{
cur = ggml_mul_mat(ctx0,
ggml_transpose(ctx0, model.layers[il].c_attn_proj_w),
cur);
cur = ggml_add(ctx0,
ggml_repeat(ctx0, model.layers[il].c_attn_proj_b, cur),
cur);
}
// add the input
cur = ggml_add(ctx0, cur, inpL);
struct ggml_tensor * inpFF = cur;
// feed-forward network
{
// norm
{
cur = ggml_norm(ctx0, inpFF);
// cur = ln_2_g*cur + ln_2_b
// [ 768, N]
cur = ggml_add(ctx0,
ggml_mul(ctx0,
ggml_repeat(ctx0, model.layers[il].ln_2_g, cur),
cur),
ggml_repeat(ctx0, model.layers[il].ln_2_b, cur));
}
// fully connected
// [3072, 768] - model.layers[il].c_mlp_fc_w
// [3072, 1] - model.layers[il].c_mlp_fc_b
// [ 768, N] - cur (in)
// [3072, N] - cur (out)
//
// cur = fc_w*cur + fc_b
// [3072, N]
cur = ggml_mul_mat(ctx0,
ggml_transpose(ctx0, model.layers[il].c_mlp_fc_w),
cur);
cur = ggml_add(ctx0,
ggml_repeat(ctx0, model.layers[il].c_mlp_fc_b, cur),
cur);
// GELU activation
// [3072, N]
cur = ggml_gelu(ctx0, cur);
// projection
// [ 768, 3072] - model.layers[il].c_mlp_proj_w
// [ 768, 1] - model.layers[il].c_mlp_proj_b
// [3072, N] - cur (in)
// [ 768, N] - cur (out)
//
// cur = proj_w*cur + proj_b
// [768, N]
cur = ggml_mul_mat(ctx0,
model.layers[il].c_mlp_proj_w_trans,
cur);
cur = ggml_add(ctx0,
ggml_repeat(ctx0, model.layers[il].c_mlp_proj_b, cur),
cur);
}
// input for next layer
inpL = ggml_add(ctx0, cur, inpFF);
}
// norm
{
// [ 768, N]
inpL = ggml_norm(ctx0, inpL);
// inpL = ln_f_g*inpL + ln_f_b
// [ 768, N]
inpL = ggml_add(ctx0,
ggml_mul(ctx0,
ggml_repeat(ctx0, model.ln_f_g, inpL),
inpL),
ggml_repeat(ctx0, model.ln_f_b, inpL));
}
// inpL = WTE * inpL
// [ 768, 50257] - model.wte
// [ 768, N] - inpL
inpL = ggml_mul_mat(ctx0, model.wte, inpL);
// logits -> probs
inpL = ggml_soft_max(ctx0, inpL);
// run the computation
ggml_build_forward_expand(&gf, inpL);
ggml_graph_compute (ctx0, &gf);
//if (n_past%100 == 0) {
// ggml_graph_print (&gf);
// ggml_graph_dump_dot(&gf, NULL, "gpt-2.dot");
//}
//embd_w.resize(n_vocab*N);
//memcpy(embd_w.data(), ggml_get_data(inpL), sizeof(float)*n_vocab*N);
// return result for just the last token
embd_w.resize(n_vocab);
memcpy(embd_w.data(), (float *) ggml_get_data(inpL) + (n_vocab*(N-1)), sizeof(float)*n_vocab);
if (mem_per_token == 0) {
mem_per_token = ggml_used_mem(ctx0)/N;
}
//printf("used_mem = %zu\n", ggml_used_mem(ctx0));
ggml_free(ctx0);
return true;
}
int main(int argc, char ** argv) {
const int64_t t_main_start_us = ggml_time_us();
gpt_params params;
params.model = "models/gpt-2-117M/ggml-model.bin";
if (gpt_params_parse(argc, argv, params) == false) {
return 1;
}
if (params.seed < 0) {
params.seed = time(NULL);
}
printf("%s: seed = %d\n", __func__, params.seed);
std::mt19937 rng(params.seed);
if (params.prompt.empty()) {
params.prompt = gpt_random_prompt(rng);
}
int64_t t_load_us = 0;
gpt_vocab vocab;
gpt2_model model;
// load the model
{
const int64_t t_start_us = ggml_time_us();
if (!gpt2_model_load(params.model, model, vocab)) {
fprintf(stderr, "%s: failed to load model from '%s'\n", __func__, params.model.c_str());
return 1;
}
t_load_us = ggml_time_us() - t_start_us;
}
int n_past = 0;
int64_t t_sample_us = 0;
int64_t t_predict_us = 0;
std::vector<float> embd_w;
// tokenize the prompt
std::vector<gpt_vocab::id> embd_inp = ::gpt_tokenize(vocab, params.prompt);
params.n_predict = std::min(params.n_predict, model.hparams.n_ctx - (int) embd_inp.size());
printf("%s: number of tokens in prompt = %zu\n", __func__, embd_inp.size());
printf("\n");
// submit the input prompt token-by-token
// this reduces the memory usage during inference, at the cost of a bit of speed at the beginning
std::vector<gpt_vocab::id> embd;
// determine the required inference memory per token:
size_t mem_per_token = 0;
gpt2_eval(model, params.n_threads, 0, { 0, 1, 2, 3 }, embd_w, mem_per_token);
for (int i = embd.size(); i < embd_inp.size() + params.n_predict; i++) {
// predict
if (embd.size() > 0) {
const int64_t t_start_us = ggml_time_us();
if (!gpt2_eval(model, params.n_threads, n_past, embd, embd_w, mem_per_token)) {
printf("Failed to predict\n");
return 1;
}
t_predict_us += ggml_time_us() - t_start_us;
}
n_past += embd.size();
embd.clear();
if (i >= embd_inp.size()) {
// sample next token
const int top_k = params.top_k;
const float top_p = params.top_p;
const float temp = params.temp;
const int n_vocab = model.hparams.n_vocab;
gpt_vocab::id id = 0;
{
const int64_t t_start_sample_us = ggml_time_us();
id = gpt_sample_top_k_top_p(vocab, embd_w.data() + (embd_w.size() - n_vocab), top_k, top_p, temp, rng);
t_sample_us += ggml_time_us() - t_start_sample_us;
}
// add it to the context
embd.push_back(id);
} else {
// if here, it means we are still processing the input prompt
for (int k = i; k < embd_inp.size(); k++) {
embd.push_back(embd_inp[k]);
if (embd.size() > params.n_batch) {
break;
}
}
i += embd.size() - 1;
}
// display text
for (auto id : embd) {
printf("%s", vocab.id_to_token[id].c_str());
}
fflush(stdout);
// end of text token
if (embd.back() == 50256) {
break;
}
}
// report timing
{
const int64_t t_main_end_us = ggml_time_us();
printf("\n\n");
printf("%s: mem per token = %8zu bytes\n", __func__, mem_per_token);
printf("%s: load time = %8.2f ms\n", __func__, t_load_us/1000.0f);
printf("%s: sample time = %8.2f ms\n", __func__, t_sample_us/1000.0f);
printf("%s: predict time = %8.2f ms / %.2f ms per token\n", __func__, t_predict_us/1000.0f, t_predict_us/1000.0f/n_past);
printf("%s: total time = %8.2f ms\n", __func__, (t_main_end_us - t_main_start_us)/1000.0f);
}
ggml_free(model.ctx);
return 0;
}

@ -0,0 +1,564 @@
<!doctype html>
<html lang="en-us">
<head>
<title>talk : GPT-2 meets Whisper in WebAssembly</title>
<style>
#output {
width: 100%;
height: 100%;
margin: 0 auto;
margin-top: 10px;
border-left: 0px;
border-right: 0px;
padding-left: 0px;
padding-right: 0px;
display: block;
background-color: black;
color: white;
font-size: 10px;
font-family: 'Lucida Console', Monaco, monospace;
outline: none;
white-space: pre;
overflow-wrap: normal;
overflow-x: scroll;
}
</style>
</head>
<body>
<div id="main-container">
<b>talk : GPT-2 meets Whisper in WebAssembly</b>
<br><br>
WIP IN PROGRESS
<br><br><hr>
<div id="model-whisper">
Whisper Model:
<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>
<!--
<input type="file" id="file" name="file" onchange="loadFile(event, 'whisper.bin')" />
-->
</div>
<br>
<div id="model-gpt-2">
GPT-2 Model:
<button id="fetch-gpt-2-small" onclick="loadGPT2('small')">small 117M (240 MB)</button>
<!--<button id="fetch-gpt-2-medium" onclick="loadGPT2('medium')">medium 345M (720 MB)</button>-->
<span id="fetch-gpt-2-progress"></span>
<!--
<input type="file" id="file" name="file" onchange="loadFile(event, 'gpt-2.bin')" />
-->
</div>
<br>
<div id="input_mic">
<button id="start" onclick="onStart()">Start</button>
<button id="stop" onclick="onStop()" disabled>Stop</button>
<button id="speak" onclick="speakTest()">Speak</button>
<button id="speak" onclick="clearCache()">Clear Cache</button>
</div>
<audio controls="controls" id="audio" loop hidden>
Your browser does not support the &lt;audio&gt; tag.
<source id="source" src="" type="audio/wav" />
</audio>
<hr><br>
<br>
<!-- textarea with height filling the rest of the page -->
<textarea id="output" rows="20"></textarea>
<br><br>
<div class="cell-version">
<span>
|
Build time: <span class="nav-link">@GIT_DATE@</span> |
Commit hash: <a class="nav-link" href="https://github.com/ggerganov/whisper.cpp/commit/@GIT_SHA1@">@GIT_SHA1@</a> |
Commit subject: <span class="nav-link">@GIT_COMMIT_SUBJECT@</span> |
<a class="nav-link" href="https://github.com/ggerganov/whisper.cpp/tree/master/examples/talk">Source Code</a> |
</span>
</div>
</div>
<script type='text/javascript'>
// TODO: convert audio buffer to WAV
function setAudio(audio) {
//if (audio) {
// // convert to 16-bit PCM
// var blob = new Blob([audio], { type: 'audio/wav' });
// var url = URL.createObjectURL(blob);
// document.getElementById('source').src = url;
// document.getElementById('audio').hidden = false;
// document.getElementById('audio').loop = false;
// document.getElementById('audio').load();
//} else {
// document.getElementById('audio').hidden = true;
//}
}
function changeInput(input) {
if (input == 'file') {
document.getElementById('input_file').style.display = 'block';
document.getElementById('input_mic').style.display = 'none';
} else {
document.getElementById('input_file').style.display = 'none';
document.getElementById('input_mic').style.display = 'block';
}
}
var printTextarea = (function() {
var element = document.getElementById('output');
if (element) element.alue = ''; // clear browser cache
return function(text) {
if (arguments.length > 1) text = Array.prototype.slice.call(arguments).join(' ');
console.log(text);
if (element) {
element.value += text + "\n";
element.scrollTop = element.scrollHeight; // focus on bottom
}
};
})();
var Module = {
print: printTextarea,
printErr: printTextarea,
setStatus: function(text) {
printTextarea('js: ' + text);
},
monitorRunDependencies: function(left) {
}
};
const kMaxAudio_s = 10;
const kRestartRecording_s = 15;
const kSampleRate = 16000;
window.AudioContext = window.AudioContext || window.webkitAudioContext;
window.OfflineAudioContext = window.OfflineAudioContext || window.webkitOfflineAudioContext;
// web audio context
var context = null;
// audio data
var audio = null;
var audio0 = null;
// the talk instance
var instance = null;
// speech synthesis
const synth = window.speechSynthesis;
// helper function
function convertTypedArray(src, type) {
var buffer = new ArrayBuffer(src.byteLength);
var baseView = new src.constructor(buffer).set(src);
return new type(buffer);
}
//
// fetch models
//
function storeFS(fname, buf) {
// write to WASM file using FS_createDataFile
// if the file exists, delete it
try {
Module.FS_unlink(fname);
} catch (e) {
// ignore
}
Module.FS_createDataFile("/", fname, buf, true, true);
printTextarea('js: stored model: ' + fname + ' size: ' + buf.length);
}
let dbVersion = 1
let dbName = 'talk.ggerganov.com';
let indexedDB = window.indexedDB || window.mozIndexedDB || window.webkitIndexedDB || window.msIndexedDB
// fetch a remote file from remote URL using the Fetch API
async function fetchRemote(url, elProgress) {
printTextarea('js: downloading with fetch()...');
const response = await fetch(
url,
{
method: 'GET',
headers: {
'Content-Type': 'application/octet-stream',
},
}
);
if (!response.ok) {
printTextarea('js: failed to fetch ' + url);
return;
}
const contentLength = response.headers.get('content-length');
const total = parseInt(contentLength, 10);
const reader = response.body.getReader();
var chunks = [];
var receivedLength = 0;
var progressLast = -1;
while (true) {
const { done, value } = await reader.read();
if (done) {
break;
}
chunks.push(value);
receivedLength += value.length;
if (contentLength) {
// update progress bar element with the new percentage
var progressCur = Math.round((receivedLength / total) * 10);
if (progressCur != progressLast) {
elProgress.innerHTML = 10*progressCur + '%';
printTextarea('js: fetching ' + 10*progressCur + '% ...');
progressLast = progressCur;
}
}
}
var chunksAll = new Uint8Array(receivedLength);
var position = 0;
for (var chunk of chunks) {
chunksAll.set(chunk, position);
position += chunk.length;
}
return chunksAll;
}
// load remote data
// - check if the data is already in the IndexedDB
// - if not, fetch it from the remote URL and store it in the IndexedDB
// - store it in WASM memory
function loadRemote(url, dst, elProgress) {
// query the storage quota and print it
navigator.storage.estimate().then(function (estimate) {
printTextarea('js: storage quota: ' + estimate.quota + ' bytes');
printTextarea('js: storage usage: ' + estimate.usage + ' bytes');
});
// check if the data is already in the IndexedDB
var request = indexedDB.open(dbName, dbVersion);
request.onupgradeneeded = function (event) {
var db = event.target.result;
if (db.version == 1) {
var objectStore = db.createObjectStore('models', { autoIncrement: false });
printTextarea('js: created IndexedDB ' + db.name + ' version ' + db.version);
} else {
// clear the database
var objectStore = event.currentTarget.transaction.objectStore('models');
objectStore.clear();
printTextarea('js: cleared IndexedDB ' + db.name + ' version ' + db.version);
}
};
request.onsuccess = function (event) {
var db = event.target.result;
var transaction = db.transaction(['models'], 'readonly');
var objectStore = transaction.objectStore('models');
var request = objectStore.get(url);
request.onsuccess = function (event) {
if (request.result) {
printTextarea('js: "' + url + '" is already in the IndexedDB');
storeFS(dst, request.result);
} else {
// data is not in the IndexedDB
printTextarea('js: "' + url + '" is not in the IndexedDB');
fetchRemote(url, elProgress).then(function (data) {
if (data) {
// store the data in the IndexedDB
var request = indexedDB.open(dbName, dbVersion);
request.onsuccess = function (event) {
var db = event.target.result;
var transaction = db.transaction(['models'], 'readwrite');
var objectStore = transaction.objectStore('models');
var request = objectStore.put(data, url);
request.onsuccess = function (event) {
printTextarea('js: "' + url + '" stored in the IndexedDB');
storeFS(dst, data);
};
request.onerror = function (event) {
printTextarea('js: failed to store "' + url + '" in the IndexedDB');
};
};
}
});
}
};
request.onerror = function (event) {
printTextarea('js: failed to get data from the IndexedDB');
};
};
request.onerror = function (event) {
printTextarea('js: failed to open IndexedDB');
};
request.onblocked = function (event) {
printTextarea('js: failed to open IndexedDB: blocked');
};
request.onabort = function (event) {
printTextarea('js: failed to open IndexedDB: abort');
};
}
function loadWhisper(model) {
let urls = {
'tiny.en': 'https://talk.ggerganov.com/ggml-model-whisper-tiny.en.bin',
'base.en': 'https://talk.ggerganov.com/ggml-model-whisper-base.en.bin',
};
let url = urls[model];
let dst = 'whisper.bin';
let el = document.getElementById('fetch-whisper-progress');
loadRemote(url, dst, el);
}
function loadGPT2(model) {
let urls = {
'small': 'https://talk.ggerganov.com/ggml-model-gpt-2-117M.bin',
'medium': 'https://talk.ggerganov.com/ggml-model-gpt-2-345M.bin',
};
let url = urls[model];
let dst = 'gpt-2.bin';
let el = document.getElementById('fetch-gpt-2-progress');
loadRemote(url, dst, el);
}
//
// microphone
//
var mediaRecorder = null;
var doRecording = false;
var startTime = 0;
function stopRecording() {
doRecording = false;
audio0 = null;
audio = null;
}
function startRecording() {
if (!context) {
context = new AudioContext({sampleRate: 16000});
}
document.getElementById('start').disabled = true;
document.getElementById('stop').disabled = false;
doRecording = true;
startTime = Date.now();
var chunks = [];
var stream = null;
navigator.mediaDevices.getUserMedia({audio: true, video: false})
.then(function(s) {
stream = s;
mediaRecorder = new MediaRecorder(stream);
mediaRecorder.ondataavailable = function(e) {
chunks.push(e.data);
var blob = new Blob(chunks, { 'type' : 'audio/ogg; codecs=opus' });
var reader = new FileReader();
reader.onload = function(event) {
var buf = new Uint8Array(reader.result);
context.decodeAudioData(buf.buffer, function(audioBuffer) {
var offlineContext = new OfflineAudioContext(audioBuffer.numberOfChannels, audioBuffer.length, audioBuffer.sampleRate);
var source = offlineContext.createBufferSource();
source.buffer = audioBuffer;
source.connect(offlineContext.destination);
source.start(0);
offlineContext.startRendering().then(function(renderedBuffer) {
audio = renderedBuffer.getChannelData(0);
//printTextarea('js: audio recorded, size: ' + audio.length + ', old size: ' + (audio0 == null ? 0 : audio0.length));
var audioAll = new Float32Array(audio0 == null ? audio.length : audio0.length + audio.length);
if (audio0 != null) {
audioAll.set(audio0, 0);
}
audioAll.set(audio, audio0 == null ? 0 : audio0.length);
if (instance) {
Module.set_audio(instance, audioAll);
}
setAudio(audio);
});
}, function(e) {
printTextarea('js: error decoding audio: ' + e);
audio = null;
setAudio(audio);
});
}
reader.readAsArrayBuffer(blob);
};
mediaRecorder.onstop = function(e) {
if (doRecording) {
setTimeout(function() {
startRecording();
});
}
};
mediaRecorder.start(250);
})
.catch(function(err) {
printTextarea('js: error getting audio stream: ' + err);
});
var interval = setInterval(function() {
if (!doRecording) {
clearInterval(interval);
mediaRecorder.stop();
stream.getTracks().forEach(function(track) {
track.stop();
});
document.getElementById('start').disabled = false;
document.getElementById('stop').disabled = true;
mediaRecorder = null;
}
// if audio length is more than kRestartRecording_s seconds, restart recording
if (audio != null && audio.length > kSampleRate*kRestartRecording_s) {
if (doRecording) {
//printTextarea('js: restarting recording');
clearInterval(interval);
audio0 = audio;
audio = null;
mediaRecorder.stop();
stream.getTracks().forEach(function(track) {
track.stop();
});
}
}
}, 250);
}
//
// speak
//
var voice = null;
function onSpeak(text) {
var voices = synth.getVoices();
var msg = new SpeechSynthesisUtterance(text);
if (voice == null) {
//voice = voices[Math.floor(Math.random() * 2)];
voice = voices[0];
}
msg.voice = voice;
synth.speak(msg);
if (doRecording) {
printTextarea('js: speaking');
stopRecording();
var interval = setInterval(function() {
if (!synth.speaking) {
printTextarea('js: done speaking');
clearInterval(interval);
startRecording();
}
}, 100);
}
}
async function clearCache() {
if (confirm('Are you sure you want to clear the cache?\nAll the models will be downloaded again.')) {
const dbs = await window.indexedDB.databases();
dbs.forEach(db => { window.indexedDB.deleteDatabase(db.name) });
}
}
//
// main
//
var intervalSpeak = null;
function onStart() {
if (!instance) {
instance = Module.init('whisper.bin');
if (instance) {
printTextarea("js: whisper initialized, instance: " + instance);
}
}
if (!instance) {
printTextarea("js: failed to initialize whisper");
return;
}
startRecording();
intervalSpeak = setInterval(function() {
var textToSpeak = Module.get_text_to_speak();
if (textToSpeak != null && textToSpeak.length > 1) {
onSpeak(textToSpeak);
}
}, 100);
}
function onStop() {
stopRecording();
if (instance) {
Module.free(instance);
instance = null;
printTextarea("js: talk instance freed");
}
}
</script>
<script type="text/javascript" src="talk.js"></script>
</body>
</html>
Loading…
Cancel
Save