vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-04 16:39:23 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

add args -k and --force

Browse source
This commit is contained in:
Zonghang Li 2025-03-11 20:44:36 +04:00
parent 9cbdf01645
commit bcfdace59b
5 changed files with 27 additions and 5 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

14
common/arg.cpp
View file

@ -737,6 +737,20 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
params.gpu_mem = value; // in GiB
}
).set_env("LLAMA_ARG_CUDA_MEM"));
add_opt(llama_arg(
{"-k", "--n-cycles"}, "N",
format("number of cycles to output one token (default: %d)", params.n_cycles),
[](gpt_params & params, int value) {
params.n_cycles = value;
}
).set_env("LLAMA_ARG_K"));
add_opt(llama_arg(
{"--force"},
format("force to start prefetching after computation (default: %s)", params.force ? "true" : "false"),
[](gpt_params & params) {
params.force = true;
}
).set_env("LLAMA_ARG_FORCE"));
// #ifdef GGML_USE_METAL
// // warn: if the output layer weights are not kept in metal shared memory, its mmap-ed weight data
// // could be released by the OS and reloaded repeatedly, which causes additional disk I/O latency.

4
common/common.cpp
View file

@ -1012,7 +1012,7 @@ static bool assign_layers_to_device(
};
// get valid factors
std::vector<int> valid_k = find_factors(n_layer);
std::vector<int> valid_k = cparams.n_cycles > 0 ? {(int)cparams.n_cycles} : find_factors(n_layer);
// assign devices to sets M1, M2, M3, and M4
// M1: devices running on macOS without Metal, and with insufficient memory
@ -1801,8 +1801,10 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
cparams.n_world = params.n_world;
cparams.rank = params.rank;
cparams.prefetch = params.prefetch;
cparams.force = params.force;
cparams.keep_out_in_metal = params.keep_out_in_metal;
cparams.n_gpu_layers = params.n_gpu_layers;
cparams.n_cycles = params.n_cycles;
std::copy(std::begin(params.n_layer_window), std::end(params.n_layer_window), cparams.n_layer_window);
if (cparams.master_ip != nullptr) {

2
common/common.h
View file

@ -149,7 +149,9 @@ struct gpt_params {
std::string next_node_ip = "localhost"; // ip address of my next node
bool prefetch = false; // prefetch layer weights
bool keep_out_in_metal = true; // whether to keep output weights in metal memory, true by default
bool force = false; // force to start prefetching after computation
int32_t gpu_mem = 999.0; // gpu memory to use, in GiB
int32_t n_cycles = 0; // number of cycles to output one token
int32_t n_predict = -1; // new tokens to predict
int32_t n_ctx = 0; // context size
int32_t n_batch = 2048; // logical batch size for prompt processing (must be >=32 to use BLAS)

2
include/llama.h
View file

@ -324,7 +324,9 @@ extern "C" {
uint32_t rank; // my rank
uint32_t n_layer_window[32];// number of layers to process in each compute
uint32_t n_gpu_layers; // number of layers to process on GPU
uint32_t n_cycles; // number of cycles to output one token
bool prefetch; // whether to prefetch layer weights
bool force; // force to start prefetching after computation
bool keep_out_in_metal; // whether to keep output weights in metal memory
char * master_ip; // ip address of the master node
char * next_node_ip; // ip address of the next node

10
src/llama.cpp
View file

@ -17892,7 +17892,7 @@ static float is_graph_loaded(struct ggml_cgraph * cgraph) {
return float(n_loaded) / float(n_total) * 100.0f;
}
static void manage_graph_tensors(struct ggml_cgraph * cgraph, int advice) {
static void manage_graph_tensors(struct ggml_cgraph * cgraph, int advice, bool force) {
long page_size = sysconf(_SC_PAGESIZE);
struct Segment {
@ -17949,7 +17949,7 @@ static void manage_graph_tensors(struct ggml_cgraph * cgraph, int advice) {
size_t len = std::max(segment.end - segment.start, static_cast<size_t>(page_size));
posix_madvise(reinterpret_cast<void *>(segment.start), len, advice); // hint to load into memory
// force to prefetch data, disabled by default
if (advice == POSIX_MADV_WILLNEED && false) {
if (advice == POSIX_MADV_WILLNEED && force) {
volatile char * ptr = reinterpret_cast<volatile char *>(segment.start);
for (size_t off = 0; off < len; off += prefetch_dense * page_size) {
for (size_t i = 0; i < prefetch_dense; i++) {
@ -18230,9 +18230,9 @@ static int llama_decode_internal(
timer(manage_graph_tensors);
int next_gf_id = (i + 1) % gf.size();
manage_graph_tensors(gf[next_gf_id], POSIX_MADV_WILLNEED);
manage_graph_tensors(gf[next_gf_id], POSIX_MADV_WILLNEED, cparams.force);
if (my_rank == 0 && (is_last_l || (next_gf_id == (int)gf.size() - 1))) {
manage_graph_tensors(gf[0], POSIX_MADV_WILLNEED);
manage_graph_tensors(gf[0], POSIX_MADV_WILLNEED, cparams.force);
}
}
}
@ -19955,7 +19955,9 @@ struct llama_context_params llama_context_default_params() {
/*.rank =*/ 0,
/*.n_layer_window =*/ {32},
/*.n_gpu_layers =*/ 0,
/*.n_cycles =*/ 0,
/*.prefetch =*/ false,
/*.force =*/ false,
/*.keep_out_in_metal =*/ true,
/*.master_ip =*/ nullptr,
/*.next_node_ip =*/ nullptr,