Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Refactor multi-thread quantize #5830

Merged
merged 1 commit into from
Mar 2, 2024
Merged

Refactor multi-thread quantize #5830

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
81 changes: 43 additions & 38 deletions llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -10819,7 +10819,7 @@ struct quantize_state_internal {
{}
};

static void llama_convert_tensor_internal(
static void llama_tensor_dequantize_internal(
struct ggml_tensor * tensor, std::vector<no_init<float>> & output, std::vector<std::thread> & workers,
const size_t nelements, const int nthread
) {
Expand Down Expand Up @@ -11160,6 +11160,46 @@ static ggml_type get_k_quant_type(quantize_state_internal & qs, ggml_type new_ty
return new_type;
}

static int32_t llama_tensor_quantize_internal(enum ggml_type new_type, const float * f32_data, void * new_data, const int chunk_size, int nrows, int n_per_row, int64_t * hist_cur, const float * imatrix, std::vector<std::thread> & workers, const int nthread) {
std::mutex mutex;
int counter = 0;
size_t new_size = 0;
if (nthread < 2) {
// single-thread
return ggml_quantize_chunk(new_type, f32_data, new_data, 0, nrows, n_per_row, hist_cur, imatrix);
}
auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, chunk_size,
nrows, n_per_row, imatrix]() {
std::array<int64_t, 1 << 4> local_hist = {};
const int nrows_per_chunk = chunk_size / n_per_row;
size_t local_size = 0;
while (true) {
std::unique_lock<std::mutex> lock(mutex);
int first_row = counter; counter += nrows_per_chunk;
if (first_row >= nrows) {
if (local_size > 0) {
for (int j=0; j<int(local_hist.size()); ++j) {
hist_cur[j] += local_hist[j];
}
new_size += local_size;
}
break;
}
lock.unlock();
const int this_nrow = std::min(nrows - first_row, nrows_per_chunk);
local_size += ggml_quantize_chunk(new_type, f32_data, new_data,
first_row * n_per_row, this_nrow, n_per_row, local_hist.data(), imatrix);
}
};
for (int it = 0; it < nthread - 1; ++it) {
workers.emplace_back(compute);
}
compute();
for (auto & w : workers) { w.join(); }
workers.clear();
return new_size;
}

static void llama_model_quantize_internal(const std::string & fname_inp, const std::string & fname_out, const llama_model_quantize_params * params) {
ggml_type quantized_type;
llama_ftype ftype = params->ftype;
Expand Down Expand Up @@ -11272,7 +11312,6 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s

std::vector<std::thread> workers;
workers.reserve(nthread);
std::mutex mutex;

int idx = 0;

Expand Down Expand Up @@ -11386,7 +11425,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
} else if (ggml_is_quantized(tensor->type) && !params->allow_requantize) {
throw std::runtime_error(format("requantizing from type %s is disabled", ggml_type_name(tensor->type)));
} else {
llama_convert_tensor_internal(tensor, f32_conv_buf, workers, nelements, nthread);
llama_tensor_dequantize_internal(tensor, f32_conv_buf, workers, nelements, nthread);
f32_data = (float *) f32_conv_buf.data();
}

Expand All @@ -11407,41 +11446,7 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s

const int nchunk = (nelements + chunk_size - 1)/chunk_size;
const int nthread_use = nthread > 1 ? std::max(1, std::min(nthread, nchunk)) : 1;
if (nthread_use < 2) {
new_size = ggml_quantize_chunk(new_type, f32_data, new_data, 0, nrows, n_per_row, hist_cur.data(), imatrix);
} else {
int counter = 0;
new_size = 0;
auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, chunk_size,
nrows, n_per_row, imatrix]() {
std::array<int64_t, 1 << 4> local_hist = {};
const int nrows_per_chunk = chunk_size / n_per_row;
size_t local_size = 0;
while (true) {
std::unique_lock<std::mutex> lock(mutex);
int first_row = counter; counter += nrows_per_chunk;
if (first_row >= nrows) {
if (local_size > 0) {
for (int j=0; j<int(local_hist.size()); ++j) {
hist_cur[j] += local_hist[j];
}
new_size += local_size;
}
break;
}
lock.unlock();
const int this_nrow = std::min(nrows - first_row, nrows_per_chunk);
local_size += ggml_quantize_chunk(new_type, f32_data, new_data,
first_row * n_per_row, this_nrow, n_per_row, local_hist.data(), imatrix);
}
};
for (int it = 0; it < nthread_use - 1; ++it) {
workers.emplace_back(compute);
}
compute();
for (auto & w : workers) { w.join(); }
workers.clear();
}
new_size = llama_tensor_quantize_internal(new_type, f32_data, new_data, chunk_size, nrows, n_per_row, hist_cur.data(), imatrix, workers, nthread_use);

LLAMA_LOG_INFO("size = %8.2f MiB -> %8.2f MiB", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
int64_t tot_count = 0;
Expand Down
Loading