public_git/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/whisper.cpp.git synced 2026-10-04 21:41:24 +02:00
Code Issues Packages Projects Releases Wiki Activity

talk-llama : sync llama.cpp

Browse Source
This commit is contained in:
Georgi Gerganov
2026-01-14 09:11:59 +02:00
parent 13dc9a912b
commit ecfcc65fbf
43 changed files with 3208 additions and 668 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/talk-llama/llama.cpp
+101 -57
View File
@@ -111,8 +111,20 @@ static std::vector<llama_device_memory_data> llama_get_device_memory_data(
}
}
for (size_t i = 0; i < ret.size(); i++) {
size_t free, total;
size_t free;
size_t total;
ggml_backend_dev_memory(model->devices[i], &free, &total);
// devices can return 0 bytes for free and total memory if they do not
// have any to report. in this case, we will use the host memory as a fallback
// fixes: https://github.com/ggml-org/llama.cpp/issues/18577
if (free == 0 && total == 0) {
ggml_backend_dev_t cpu_dev = ggml_backend_dev_by_type(GGML_BACKEND_DEVICE_TYPE_CPU);
if (cpu_dev == nullptr) {
throw std::runtime_error(format("%s: no CPU backend found", __func__));
}
ggml_backend_dev_memory(cpu_dev, &free, &total);
}
ret[i].free = free;
ret[i].total = total;
}
@@ -147,9 +159,8 @@ class llama_params_fit_exception : public std::runtime_error {
static void llama_params_fit_impl(
const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
size_t margin_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
size_t * margins_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
constexpr int64_t MiB = 1024*1024;
const int64_t margin = margin_s; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
typedef std::vector<llama_device_memory_data> dmds_t;
const llama_model_params default_mparams = llama_model_default_params();
@@ -168,6 +179,12 @@ static void llama_params_fit_impl(
return;
}
std::vector<int64_t> margins; // this function uses int64_t rather than size_t for memory sizes to more conveniently handle deficits
margins.reserve(nd);
for (size_t id = 0; id < nd; id++) {
margins.push_back(margins_s[id]);
}
std::vector<std::string> dev_names;
{
dev_names.reserve(nd);
@@ -187,9 +204,10 @@ static void llama_params_fit_impl(
int64_t sum_free = 0;
int64_t sum_projected_free = 0;
int64_t min_projected_free = INT64_MAX;
int64_t sum_projected_used = 0;
int64_t sum_projected_model = 0;
std::vector<int64_t> projected_free_per_device;
projected_free_per_device.reserve(nd);
if (nd > 1) {
LLAMA_LOG_INFO("%s: projected memory use with initial parameters [MiB]:\n", __func__);
@@ -199,45 +217,63 @@ static void llama_params_fit_impl(
const int64_t projected_used = dmd.mb.total();
const int64_t projected_free = dmd.free - projected_used;
projected_free_per_device.push_back(projected_free);
sum_free += dmd.free;
sum_projected_used += projected_used;
sum_projected_free += projected_free;
min_projected_free = std::min(min_projected_free, projected_free);
sum_projected_model += dmd.mb.model;
if (nd > 1) {
LLAMA_LOG_INFO("%s: - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " %s\n",
__func__, dev_names[id].c_str(), dmd.total/MiB, projected_used/MiB, std::abs(projected_free)/MiB,
projected_free >= 0 ? "surplus" : "deficit");
LLAMA_LOG_INFO("%s: - %s: %6" PRId64 " total, %6" PRId64 " used, %6" PRId64 " free vs. target of %6" PRId64 "\n",
__func__, dev_names[id].c_str(), dmd.total/MiB, projected_used/MiB, projected_free/MiB, margins[id]/MiB);
}
}
assert(sum_free >= 0 && sum_projected_used >= 0);
LLAMA_LOG_INFO("%s: projected to use %" PRId64 " MiB of device memory vs. %" PRId64 " MiB of free device memory\n",
__func__, sum_projected_used/MiB, sum_free/MiB);
if (min_projected_free >= margin) {
if (nd == 1) {
if (nd == 1) {
if (projected_free_per_device[0] >= margins[0]) {
LLAMA_LOG_INFO("%s: will leave %" PRId64 " >= %" PRId64 " MiB of free device memory, no changes needed\n",
__func__, min_projected_free/MiB, margin/MiB);
__func__, projected_free_per_device[0]/MiB, margins[0]/MiB);
return;
}
} else {
bool changes_needed = false;
for (size_t id = 0; id < nd; id++) {
if (projected_free_per_device[id] < margins[id]) {
changes_needed = true;
break;
}
}
if (!changes_needed) {
LLAMA_LOG_INFO("%s: targets for free memory can be met on all devices, no changes needed\n", __func__);
return;
}
LLAMA_LOG_INFO("%s: will leave at least %" PRId64 " >= %" PRId64 " MiB of free memory on all devices, no changes needed\n",
__func__, min_projected_free/MiB, margin/MiB);
return;
}
// step 2: try reducing memory use by reducing the context size
{
int64_t global_surplus = sum_projected_free - int64_t(nd)*margin;
int64_t global_surplus = sum_projected_free;
for (size_t id = 0; id < nd; id++) {
global_surplus -= margins[id];
}
if (global_surplus < 0) {
LLAMA_LOG_INFO(nd == 1 ?
"%s: cannot fulfill margin of %" PRId64 " MiB, need to reduce device memory by %" PRId64 " MiB\n" :
"%s: cannot fulfill margin of %" PRId64 " MiB on all devices, need to use %" PRId64 " MiB less in total\n",
__func__, margin/MiB, -global_surplus/MiB);
if (nd == 1) {
LLAMA_LOG_INFO("%s: cannot meet free memory target of %" PRId64 " MiB, need to reduce device memory by %" PRId64 " MiB\n",
__func__, margins[0]/MiB, -global_surplus/MiB);
} else {
LLAMA_LOG_INFO(
"%s: cannot meet free memory targets on all devices, need to use %" PRId64 " MiB less in total\n",
__func__, -global_surplus/MiB);
}
if (cparams->n_ctx == 0) {
if (hp_nct > n_ctx_min) {
int64_t sum_used_target = sum_free - nd*margin_s;
int64_t sum_used_target = sum_free;
for (size_t id = 0; id < nd; id++) {
sum_used_target -= margins[id];
}
if (nd > 1) {
// for multiple devices we need to be more conservative in terms of how much context we think can fit:
// - for dense models only whole layers can be assigned to devices
@@ -359,6 +395,11 @@ static void llama_params_fit_impl(
// for the first partial layer varying parts can overflow, all further layers use LAYER_FRACTION_MOE:
layer_fraction_t overflow_type = LAYER_FRACTION_MOE;
uint32_t n_full() const {
assert(n_layer >= n_part);
return n_layer - n_part;
}
};
const size_t ntbo = llama_max_tensor_buft_overrides();
@@ -382,7 +423,7 @@ static void llama_params_fit_impl(
size_t itbo = 0;
for (size_t id = 0; id < nd; id++) {
il0 += ngl_per_device[id].n_layer - ngl_per_device[id].n_part;
il0 += ngl_per_device[id].n_full();
for (uint32_t il = il0; il < il0 + ngl_per_device[id].n_part; il++) {
if (itbo + 1 >= ntbo) {
tensor_buft_overrides[itbo].pattern = nullptr;
@@ -393,7 +434,7 @@ static void llama_params_fit_impl(
+ std::to_string(ntbo) + " is insufficient for model");
}
tensor_buft_overrides[itbo].pattern = get_overflow_pattern(il, il == il0 ? ngl_per_device[id].overflow_type : LAYER_FRACTION_MOE);
tensor_buft_overrides[itbo].buft = overflow_bufts[id];
tensor_buft_overrides[itbo].buft = il == il0 ? overflow_bufts[id] : ggml_backend_cpu_buffer_type();
itbo++;
}
il0 += ngl_per_device[id].n_part;
@@ -443,9 +484,9 @@ static void llama_params_fit_impl(
const dmds_t dmds_cpu_moe = llama_get_device_memory_data(
path_model, mparams, cparams, devs, hp_ngl, hp_nct, hp_nex, log_level);
for (const llama_device_memory_data & dmd : dmds_cpu_moe) {
global_surplus_cpu_moe += dmd.free;
global_surplus_cpu_moe -= int64_t(dmd.mb.total()) + margin;
for (size_t id = 0; id < nd; id++) {
global_surplus_cpu_moe += dmds_cpu_moe[id].free;
global_surplus_cpu_moe -= int64_t(dmds_cpu_moe[id].mb.total()) + margins[id];
}
if (global_surplus_cpu_moe > 0) {
@@ -464,24 +505,18 @@ static void llama_params_fit_impl(
std::vector<int64_t> targets; // maximum acceptable memory use per device
targets.reserve(nd);
for (size_t id = 0; id < nd; id++) {
targets.push_back(dmds_full[id].free - margin);
targets.push_back(dmds_full[id].free - margins[id]);
LLAMA_LOG_DEBUG("%s: id=%zu, target=%" PRId64 " MiB\n", __func__, id, targets[id]/MiB);
}
std::vector<ggml_backend_buffer_type_t> overflow_bufts; // which bufts the partial layers of a device overflow to:
std::vector<ggml_backend_buffer_type_t> overflow_bufts; // which bufts the first partial layer of a device overflows to:
overflow_bufts.reserve(nd);
for (size_t id = 0; id < nd - 1; ++id) {
overflow_bufts.push_back(ggml_backend_dev_buffer_type(devs[id + 1]));
for (size_t id = 0; id < nd; id++) {
overflow_bufts.push_back(ggml_backend_cpu_buffer_type());
}
overflow_bufts.push_back(ggml_backend_cpu_buffer_type());
std::vector<ngl_t> ngl_per_device(nd);
std::vector<int64_t> mem = get_memory_for_layers(__func__, ngl_per_device, overflow_bufts);
if (hp_nex > 0) {
for (size_t id = 0; id < nd; id++) {
ngl_per_device[id].overflow_type = LAYER_FRACTION_MOE;
}
}
// optimize the number of layers per device using the method of false position:
// - ngl_per_device has 0 layers for each device, lower bound
@@ -512,9 +547,6 @@ static void llama_params_fit_impl(
if (mem_high[id] > targets[id]) {
assert(ngl_per_device_high[id].n_layer > ngl_per_device[id].n_layer);
uint32_t delta = ngl_per_device_high[id].n_layer - ngl_per_device[id].n_layer;
if (hp_nex > 0 && size_t(id) == nd - 1) {
delta--;
}
LLAMA_LOG_DEBUG("%s: start filling device %" PRIu32 ", delta=%" PRIu32 "\n", __func__, id, delta);
while (delta > 1) {
uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
@@ -524,7 +556,8 @@ static void llama_params_fit_impl(
std::vector<ngl_t> ngl_per_device_test = ngl_per_device;
ngl_per_device_test[id].n_layer += step_size;
if (hp_nex) {
ngl_per_device_test[id].n_part += step_size;
ngl_per_device_test[id].n_part += size_t(id) == nd - 1 && ngl_per_device_test[id].n_part == 0 ?
step_size - 1 : step_size; // the first layer is the output layer which must always be full
}
const std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts);
@@ -573,7 +606,7 @@ static void llama_params_fit_impl(
assert(id_dense_start < nd);
LLAMA_LOG_INFO("%s: converting dense-only layers to full layers and filling them front-to-back with overflow to next device/system memory:\n", __func__);
for (size_t id = 0; id <= id_dense_start; id++) {
for (size_t id = 0; id <= id_dense_start && id_dense_start < nd; id++) {
std::vector<ngl_t> ngl_per_device_high = ngl_per_device;
for (size_t jd = id_dense_start; jd < nd; jd++) {
const uint32_t n_layer_move = jd < nd - 1 ? ngl_per_device_high[jd].n_layer : ngl_per_device_high[jd].n_layer - 1;
@@ -585,12 +618,8 @@ static void llama_params_fit_impl(
std::vector<int64_t> mem_high = get_memory_for_layers(__func__, ngl_per_device_high, overflow_bufts);
if (mem_high[id] > targets[id]) {
assert(ngl_per_device_high[id].n_layer >= ngl_per_device_high[id].n_part);
assert(ngl_per_device[id].n_layer >= ngl_per_device[id].n_part);
assert((ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
>= ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
uint32_t delta = (ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
- (ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
assert(ngl_per_device_high[id].n_full() >= ngl_per_device[id].n_full());
uint32_t delta = ngl_per_device_high[id].n_full() - ngl_per_device[id].n_full();
while (delta > 1) {
uint32_t step_size = int64_t(delta) * (targets[id] - mem[id]) / (mem_high[id] - mem[id]);
step_size = std::max(step_size, uint32_t(1));
@@ -606,7 +635,7 @@ static void llama_params_fit_impl(
ngl_per_device_test[id].n_layer += n_convert_jd;
n_converted_test += n_convert_jd;
if (ngl_per_device_test[id_dense_start_test].n_layer > 0) {
if (ngl_per_device_test[id_dense_start_test].n_part > 0) {
break;
}
}
@@ -625,8 +654,8 @@ static void llama_params_fit_impl(
LLAMA_LOG_DEBUG("%s: set ngl_per_device_high[%zu].(n_layer, n_part)=(%" PRIu32 ", %" PRIu32 "), id_dense_start_high=%zu\n",
__func__, id, ngl_per_device_high[id].n_layer, ngl_per_device_high[id].n_part, id_dense_start_high);
}
delta = (ngl_per_device_high[id].n_layer - ngl_per_device_high[id].n_part)
- (ngl_per_device[id].n_layer - ngl_per_device[id].n_part);
assert(ngl_per_device_high[id].n_full() >= ngl_per_device[id].n_full());
delta = ngl_per_device_high[id].n_full() - ngl_per_device[id].n_full();
}
} else {
ngl_per_device = ngl_per_device_high;
@@ -644,14 +673,19 @@ static void llama_params_fit_impl(
ngl_per_device_test[id_dense_start_test].n_part--;
ngl_per_device_test[id].n_layer++;
ngl_per_device_test[id].n_part++;
if (ngl_per_device_test[id_dense_start_test].n_layer == 0) {
if (ngl_per_device_test[id_dense_start_test].n_part == 0) {
id_dense_start_test++;
}
ngl_per_device_test[id].overflow_type = LAYER_FRACTION_UP;
std::vector<ggml_backend_buffer_type_t> overflow_bufts_test = overflow_bufts;
if (id < nd - 1) {
overflow_bufts_test[id] = ggml_backend_dev_buffer_type(devs[id + 1]);
}
LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_UP\n", __func__);
std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts);
std::vector<int64_t> mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
ngl_per_device = ngl_per_device_test;
overflow_bufts = overflow_bufts_test;
mem = mem_test;
id_dense_start = id_dense_start_test;
LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", UP), id_dense_start=%zu\n",
@@ -659,9 +693,10 @@ static void llama_params_fit_impl(
ngl_per_device_test[id].overflow_type = LAYER_FRACTION_GATE;
LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_GATE\n", __func__);
mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts);
mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
ngl_per_device = ngl_per_device_test;
overflow_bufts = overflow_bufts_test;
mem = mem_test;
id_dense_start = id_dense_start_test;
LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", GATE), id_dense_start=%zu\n",
@@ -670,9 +705,10 @@ static void llama_params_fit_impl(
} else {
ngl_per_device_test[id].overflow_type = LAYER_FRACTION_ATTN;
LLAMA_LOG_DEBUG("%s: trying to fit one extra layer with overflow_type=LAYER_FRACTION_ATTN\n", __func__);
mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts);
mem_test = get_memory_for_layers(__func__, ngl_per_device_test, overflow_bufts_test);
if (mem_test[id] < targets[id] && (id + 1 == nd || mem_test[id + 1] < targets[id + 1])) {
ngl_per_device = ngl_per_device_test;
overflow_bufts = overflow_bufts_test;
mem = mem_test;
id_dense_start = id_dense_start_test;
LLAMA_LOG_DEBUG("%s: set ngl_per_device[%zu].(n_layer, n_part, overflow_type)=(%" PRIu32 ", %" PRIu32 ", ATTN), id_dense_start=%zu\n",
@@ -687,17 +723,25 @@ static void llama_params_fit_impl(
__func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, ngl_per_device[id].n_part, mem[id]/MiB, projected_margin/MiB);
}
// print info for devices that were not changed during the conversion from dense only to full layers:
for (size_t id = id_dense_start + 1; id < nd; id++) {
const int64_t projected_margin = dmds_full[id].free - mem[id];
LLAMA_LOG_INFO(
"%s: - %s: %2" PRIu32 " layers (%2" PRIu32 " overflowing), %6" PRId64 " MiB used, %6" PRId64 " MiB free\n",
__func__, dev_names[id].c_str(), ngl_per_device[id].n_layer, ngl_per_device[id].n_part, mem[id]/MiB, projected_margin/MiB);
}
set_ngl_tensor_split_tbo(ngl_per_device, overflow_bufts, *mparams);
}
enum llama_params_fit_status llama_params_fit(
const char * path_model, struct llama_model_params * mparams, struct llama_context_params * cparams,
float * tensor_split, struct llama_model_tensor_buft_override * tensor_buft_overrides,
size_t margin_s, uint32_t n_ctx_min, enum ggml_log_level log_level) {
size_t * margins, uint32_t n_ctx_min, enum ggml_log_level log_level) {
const int64_t t0_us = llama_time_us();
llama_params_fit_status status = LLAMA_PARAMS_FIT_STATUS_SUCCESS;
try {
llama_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margin_s, n_ctx_min, log_level);
llama_params_fit_impl(path_model, mparams, cparams, tensor_split, tensor_buft_overrides, margins, n_ctx_min, log_level);
LLAMA_LOG_INFO("%s: successfully fit params to free device memory\n", __func__);
} catch (const llama_params_fit_exception & e) {
LLAMA_LOG_WARN("%s: failed to fit params to free device memory: %s\n", __func__, e.what());
@@ -713,7 +757,7 @@ enum llama_params_fit_status llama_params_fit(
struct llama_sampler_chain_params llama_sampler_chain_default_params() {
struct llama_sampler_chain_params result = {
/*.no_perf =*/ true,
/*.no_perf =*/ true,
};
return result;
@@ -786,7 +830,7 @@ static int llama_model_load(const std::string & fname, std::vector<std::string>
model.t_start_us = tm.t_start_us;
try {
llama_model_loader ml(fname, splits, params.use_mmap, params.check_tensors, params.no_alloc, params.kv_overrides, params.tensor_buft_overrides);
llama_model_loader ml(fname, splits, params.use_mmap, params.use_direct_io, params.check_tensors, params.no_alloc, params.kv_overrides, params.tensor_buft_overrides);
ml.print_info();