Skip to content

llama : make quantize example up to 2.7x faster #3115

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

Merged
merged 3 commits into from
Sep 15, 2023
Merged

llama : make quantize example up to 2.7x faster #3115

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
0c64968
llama : refactor k-quant mixture logic into a function
cebtenzzre Sep 9, 2023
a95aa21
llama : optimize vector use in quantize -> 179% faster
cebtenzzre Sep 9, 2023
f727ad5
llama : don't zero-init vectors in quantize -> 5.1% faster
cebtenzzre Sep 9, 2023
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
267 changes: 142 additions & 125 deletions llama.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4639,7 +4639,16 @@ void llama_beam_search(llama_context * ctx,
// quantization
//

static void llama_convert_tensor_internal(struct ggml_tensor * tensor, std::vector<float> & output, const size_t nelements, const int nthread) {
template <typename T>
struct no_init {
T value;
no_init() { /* do nothing */ }
};

static void llama_convert_tensor_internal(
struct ggml_tensor * tensor, std::vector<no_init<float>> & output, std::vector<std::thread> & workers,
const size_t nelements, const int nthread
) {
if (output.size() < nelements) {
output.resize(nelements);
}
Expand Down Expand Up @@ -4674,7 +4683,6 @@ static void llama_convert_tensor_internal(struct ggml_tensor * tensor, std::vect
auto blocks_per_thread = nblocks / nthread;
auto spare_blocks = nblocks - (blocks_per_thread * nthread); // if blocks aren't divisible by thread count

std::vector<std::thread> workers;
for (auto tnum = 0, in_buff_offs = 0, out_buff_offs = 0; tnum < nthread; tnum++) {
auto thr_blocks = blocks_per_thread + (tnum == nthread - 1 ? spare_blocks : 0); // num blocks for this thread
auto thr_elems = thr_blocks * block_size; // number of elements for this thread
Expand All @@ -4687,14 +4695,123 @@ static void llama_convert_tensor_internal(struct ggml_tensor * tensor, std::vect
qtype.to_float(inbuf, outbuf, nels);
}
};
workers.push_back(std::thread(compute, tensor->type, (uint8_t *) tensor->data + in_buff_offs, f32_output + out_buff_offs, thr_elems));
workers.emplace_back(compute, tensor->type, (uint8_t *) tensor->data + in_buff_offs, f32_output + out_buff_offs, thr_elems);
in_buff_offs += thr_block_bytes;
out_buff_offs += thr_elems;
}
for (auto & worker : workers) {
worker.join();
for (auto & w : workers) { w.join(); }
workers.clear();
}

#ifdef GGML_USE_K_QUANTS
static ggml_type get_k_quant_type(
ggml_type new_type, const ggml_tensor * tensor, const llama_model & model, llama_ftype ftype, int * i_attention_wv,
int n_attention_wv, int * i_feed_forward_w2, int n_feed_forward_w2
) {
const std::string name = ggml_get_name(tensor);
// TODO: avoid hardcoded tensor names - use the TN_* constants
const auto tn = LLM_TN(model.arch);

auto use_more_bits = [](int i_layer, int num_layers) -> bool {
return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2;
};

if (name == tn(LLM_TENSOR_OUTPUT, "weight")) {
int nx = tensor->ne[0];
if (model.arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
new_type = GGML_TYPE_Q8_0;
}
else if (new_type != GGML_TYPE_Q8_0) {
new_type = GGML_TYPE_Q6_K;
}
} else if (name.find("attn_v.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
new_type = *i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
use_more_bits(*i_attention_wv, n_attention_wv)) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && *i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) &&
(*i_attention_wv < n_attention_wv/8 || *i_attention_wv >= 7*n_attention_wv/8)) new_type = GGML_TYPE_Q6_K;
if (model.type == MODEL_70B) {
// In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
// 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
// nearly negligible increase in model size by quantizing this tensor with more bits:
if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K;
}
++*i_attention_wv;
} else if (name.find("ffn_down.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
new_type = *i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K
: model.arch != LLM_ARCH_FALCON || use_more_bits(*i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q4_K
: GGML_TYPE_Q3_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
new_type = model.arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
if (model.arch == LLM_ARCH_FALCON) {
new_type = *i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K :
use_more_bits(*i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
} else {
if (use_more_bits(*i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
}
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(*i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && model.arch != LLM_ARCH_FALCON && *i_feed_forward_w2 < 4) {
new_type = GGML_TYPE_Q5_K;
}
++*i_feed_forward_w2;
} else if (name.find("attn_output.weight") != std::string::npos) {
if (model.arch != LLM_ARCH_FALCON) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
} else {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
}
}
else if (name.find("attn_qkv.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
}
else if (name.find("ffn_gate.weight") != std::string::npos || name.find("ffn_up.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
}
// This can be used to reduce the size of the Q5_K_S model.
// The associated PPL increase is fully in line with the size reduction
//else {
// if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K;
//}
bool convert_incompatible_tensor = false;
if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K ||
new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K) {
int nx = tensor->ne[0];
int ny = tensor->ne[1];
if (nx % QK_K != 0) {
LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for k-quants\n", __func__, nx, ny, QK_K);
convert_incompatible_tensor = true;
}
}
if (convert_incompatible_tensor) {
if (name == tn(LLM_TENSOR_OUTPUT, "weight")) {
new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing.
LLAMA_LOG_WARN("F16 will be used for this tensor instead.\n");
} else if (name == tn(LLM_TENSOR_TOKEN_EMBD, "weight")) {
new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing.
LLAMA_LOG_WARN("Q4_0 will be used for this tensor instead.\n");
} else {
throw std::runtime_error("Unsupported tensor size encountered\n");
}
}

return new_type;
}
#endif

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;
Expand Down Expand Up @@ -4779,18 +4896,14 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
std::vector<int64_t> hist_all(1 << 4, 0);

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

#ifdef GGML_USE_K_QUANTS
auto use_more_bits = [] (int i_layer, int num_layers) -> bool {
return i_layer < num_layers/8 || i_layer >= 7*num_layers/8 || (i_layer - num_layers/8)%3 == 2;
};
#endif

int idx = 0;

std::vector<uint8_t> read_data;
std::vector<uint8_t> work;
std::vector<no_init<uint8_t>> read_data;
std::vector<no_init<uint8_t>> work;
std::vector<no_init<float>> f32_conv_buf;

// populate the original tensors so we get an initial meta data
for (int i = 0; i < ml->n_tensors; ++i) {
Expand All @@ -4812,7 +4925,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s

const std::string name = ggml_get_name(tensor);

read_data.resize(ggml_nbytes(tensor));
if (read_data.size() < ggml_nbytes(tensor)) {
read_data.resize(ggml_nbytes(tensor));
}
tensor->data = read_data.data();
ml->load_data_for(tensor);

Expand All @@ -4837,101 +4952,9 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
if (quantize) {
new_type = quantized_type;
#ifdef GGML_USE_K_QUANTS
// TODO: avoid hardcoded tensor names - use the TN_* constants
const auto tn = LLM_TN(ml->get_arch());

if (name == tn(LLM_TENSOR_OUTPUT, "weight")) {
int nx = tensor->ne[0];
if (model.arch == LLM_ARCH_FALCON || nx % QK_K != 0) {
new_type = GGML_TYPE_Q8_0;
}
else if (new_type != GGML_TYPE_Q8_0) {
new_type = GGML_TYPE_Q6_K;
}
} else if (name.find("attn_v.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
new_type = i_attention_wv < 2 ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
else if ((ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) &&
use_more_bits(i_attention_wv, n_attention_wv)) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && i_attention_wv < 4) new_type = GGML_TYPE_Q5_K;
else if (QK_K == 64 && (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_S) &&
(i_attention_wv < n_attention_wv/8 || i_attention_wv >= 7*n_attention_wv/8)) new_type = GGML_TYPE_Q6_K;
if (model.type == MODEL_70B) {
// In the 70B model we have 8 heads sharing the same attn_v weights. As a result, the attn_v.weight tensor is
// 8x smaller compared to attn_q.weight. Hence, we can get a nice boost in quantization accuracy with
// nearly negligible increase in model size by quantizing this tensor with more bits:
if (new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K) new_type = GGML_TYPE_Q5_K;
}
++i_attention_wv;
} else if (name.find("ffn_down.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) {
new_type = i_feed_forward_w2 < 2 ? GGML_TYPE_Q5_K
: model.arch != LLM_ARCH_FALCON || use_more_bits(i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q4_K
: GGML_TYPE_Q3_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) {
new_type = model.arch == LLM_ARCH_FALCON ? GGML_TYPE_Q4_K : GGML_TYPE_Q5_K;
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) {
if (model.arch == LLM_ARCH_FALCON) {
new_type = i_feed_forward_w2 < 2 ? GGML_TYPE_Q6_K :
use_more_bits(i_feed_forward_w2, n_feed_forward_w2) ? GGML_TYPE_Q5_K : GGML_TYPE_Q4_K;
} else {
if (use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
}
}
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M && use_more_bits(i_feed_forward_w2, n_feed_forward_w2)) new_type = GGML_TYPE_Q6_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_S && model.arch != LLM_ARCH_FALCON && i_feed_forward_w2 < 4) {
new_type = GGML_TYPE_Q5_K;
}
++i_feed_forward_w2;
} else if (name.find("attn_output.weight") != std::string::npos) {
if (model.arch != LLM_ARCH_FALCON) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K ) new_type = GGML_TYPE_Q3_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M) new_type = GGML_TYPE_Q4_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q5_K;
} else {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
}
}
else if (name.find("attn_qkv.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q3_K_M || ftype == LLAMA_FTYPE_MOSTLY_Q3_K_L) new_type = GGML_TYPE_Q4_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q4_K_M) new_type = GGML_TYPE_Q5_K;
else if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_M) new_type = GGML_TYPE_Q6_K;
}
else if (name.find("ffn_gate.weight") != std::string::npos || name.find("ffn_up.weight") != std::string::npos) {
if (ftype == LLAMA_FTYPE_MOSTLY_Q2_K) new_type = GGML_TYPE_Q3_K;
}
// This can be used to reduce the size of the Q5_K_S model.
// The associated PPL increase is fully in line with the size reduction
//else {
// if (ftype == LLAMA_FTYPE_MOSTLY_Q5_K_S) new_type = GGML_TYPE_Q4_K;
//}
bool convert_incompatible_tensor = false;
if (new_type == GGML_TYPE_Q2_K || new_type == GGML_TYPE_Q3_K || new_type == GGML_TYPE_Q4_K ||
new_type == GGML_TYPE_Q5_K || new_type == GGML_TYPE_Q6_K) {
int nx = tensor->ne[0];
int ny = tensor->ne[1];
if (nx % QK_K != 0) {
LLAMA_LOG_WARN("\n\n%s : tensor cols %d x %d are not divisible by %d, required for k-quants\n", __func__, nx, ny, QK_K);
convert_incompatible_tensor = true;
}
}
if (convert_incompatible_tensor) {
if (name == tn(LLM_TENSOR_OUTPUT, "weight")) {
new_type = GGML_TYPE_F16; //fall back to F16 instead of just failing.
LLAMA_LOG_WARN("F16 will be used for this tensor instead.\n");
} else if (name == tn(LLM_TENSOR_TOKEN_EMBD, "weight")) {
new_type = GGML_TYPE_Q4_0; //fall back to Q4_0 instead of just failing.
LLAMA_LOG_WARN("Q4_0 will be used for this tensor instead.\n");
} else {
throw std::runtime_error("Unsupported tensor size encountered\n");
}
}
new_type = get_k_quant_type(
new_type, tensor, model, ftype, &i_attention_wv, n_attention_wv, &i_feed_forward_w2, n_feed_forward_w2
);
#endif
// If we've decided to quantize to the same type the tensor is already
// in then there's nothing to do.
Expand All @@ -4946,23 +4969,24 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
const size_t nelements = ggml_nelements(tensor);

float * f32_data;
std::vector<float> f32_conv_buf;

if (tensor->type == GGML_TYPE_F32) {
f32_data = (float *) tensor->data;
} 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, nelements, nthread);
llama_convert_tensor_internal(tensor, f32_conv_buf, workers, nelements, nthread);
f32_data = (float *) f32_conv_buf.data();
}

LLAMA_LOG_INFO("quantizing to %s .. ", ggml_type_name(new_type));
fflush(stdout);

work.resize(nelements * 4); // upper bound on size
if (work.size() < nelements * 4) {
work.resize(nelements * 4); // upper bound on size
}
new_data = work.data();
std::vector<int64_t> hist_cur(1 << 4, 0);
std::array<int64_t, 1 << 4> hist_cur = {};

static const int chunk_size = 32 * 512;
const int nchunk = (nelements + chunk_size - 1)/chunk_size;
Expand All @@ -4973,13 +4997,13 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
size_t counter = 0;
new_size = 0;
auto compute = [&mutex, &counter, &hist_cur, &new_size, new_type, f32_data, new_data, nelements]() {
std::vector<int64_t> local_hist;
std::array<int64_t, 1 << 4> local_hist = {};
size_t local_size = 0;
while (true) {
std::unique_lock<std::mutex> lock(mutex);
size_t first = counter; counter += chunk_size;
if (first >= nelements) {
if (!local_hist.empty()) {
if (local_size > 0) {
for (int j=0; j<int(local_hist.size()); ++j) {
hist_cur[j] += local_hist[j];
}
Expand All @@ -4989,22 +5013,15 @@ static void llama_model_quantize_internal(const std::string & fname_inp, const s
}
lock.unlock();
size_t last = std::min(nelements, first + chunk_size);
if (local_hist.empty()) {
local_hist.resize(hist_cur.size(), 0);
}
local_size += ggml_quantize_chunk(new_type, f32_data, new_data, first, last - first, local_hist.data());
}
};
if ((int) workers.size() < nthread_use - 1) {
workers.resize(nthread_use - 1);
}
for (int it = 0; it < nthread_use - 1; ++it) {
workers[it] = std::thread(compute);
workers.emplace_back(compute);
}
compute();
for (int it = 0; it < nthread_use - 1; ++it) {
workers[it].join();
}
for (auto & w : workers) { w.join(); }
workers.clear();
}

LLAMA_LOG_INFO("size = %8.2f MB -> %8.2f MB | hist: ", ggml_nbytes(tensor)/1024.0/1024.0, new_size/1024.0/1024.0);
Expand Down