vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-06 05:09:03 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

use arg prefetch and remove arg unload

Browse source
This commit is contained in:
Lizonghang 2025-02-12 17:04:41 +04:00
parent 708b1d8c89
commit c84f9d29fe
5 changed files with 15 additions and 19 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

6
common/arg.cpp
View file

@ -724,10 +724,10 @@ gpt_params_context gpt_params_parser_init(gpt_params & params, llama_example ex,
} }
).set_env("LLAMA_ARG_NEXT_NODE_IP")); ).set_env("LLAMA_ARG_NEXT_NODE_IP"));
add_opt(llama_arg( add_opt(llama_arg(
{"--unload", "--unload-weight"}, {"--prefetch"},
format("whether to unload layer weights after use (default: %s)", params.unload ? "true" : "false"), format("whether to prefetch layer weights (default: %s)", params.prefetch ? "true" : "false"),
[](gpt_params & params) { [](gpt_params & params) {
params.unload = true; params.prefetch = true;
} }
).set_env("LLAMA_ARG_UNLOAD")); ).set_env("LLAMA_ARG_UNLOAD"));
add_opt(llama_arg( add_opt(llama_arg(

2
common/common.cpp
View file

@ -1714,7 +1714,7 @@ struct llama_context_params llama_context_params_from_gpt_params(const gpt_param
cparams.n_world = params.n_world; cparams.n_world = params.n_world;
cparams.rank = params.rank; cparams.rank = params.rank;
cparams.unload = params.unload; cparams.prefetch = params.prefetch;
cparams.keep_out_in_metal = params.keep_out_in_metal; cparams.keep_out_in_metal = params.keep_out_in_metal;
cparams.n_gpu_layers = params.n_gpu_layers; cparams.n_gpu_layers = params.n_gpu_layers;
std::copy(std::begin(params.n_layer_window), std::end(params.n_layer_window), cparams.n_layer_window); std::copy(std::begin(params.n_layer_window), std::end(params.n_layer_window), cparams.n_layer_window);

2
common/common.h
View file

@ -147,7 +147,7 @@ struct gpt_params {
uint32_t n_layer_window[32] = {0}; // layer window size on each node uint32_t n_layer_window[32] = {0}; // layer window size on each node
std::string master_ip = "localhost"; // ip address of the master node std::string master_ip = "localhost"; // ip address of the master node
std::string next_node_ip = "localhost"; // ip address of my next node std::string next_node_ip = "localhost"; // ip address of my next node
bool unload = false; // unload layer weights after use or not bool prefetch = false; // prefetch layer weights
bool keep_out_in_metal = true; // whether to keep output weights in metal memory, true by default bool keep_out_in_metal = true; // whether to keep output weights in metal memory, true by default
int32_t gpu_mem = 999.0; // gpu memory to use, in GiB int32_t gpu_mem = 999.0; // gpu memory to use, in GiB
int32_t n_predict = -1; // new tokens to predict int32_t n_predict = -1; // new tokens to predict

2
include/llama.h
View file

@ -322,7 +322,7 @@ extern "C" {
uint32_t rank; // my rank uint32_t rank; // my rank
uint32_t n_layer_window[32];// number of layers to process in each compute 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_gpu_layers; // number of layers to process on GPU
bool unload; // whether to unload layer weights after use bool prefetch; // whether to prefetch layer weights
bool keep_out_in_metal; // whether to keep output weights in metal memory bool keep_out_in_metal; // whether to keep output weights in metal memory
char * master_ip; // ip address of the master node char * master_ip; // ip address of the master node
char * next_node_ip; // ip address of the next node char * next_node_ip; // ip address of the next node

22
src/llama.cpp
View file

@ -2571,7 +2571,7 @@ struct llama_cparams {
uint32_t n_world; uint32_t n_world;
uint32_t rank; uint32_t rank;
uint32_t n_layer_window[32]; uint32_t n_layer_window[32];
bool unload; bool prefetch;
uint32_t n_ctx; // context size used during inference uint32_t n_ctx; // context size used during inference
uint32_t n_batch; uint32_t n_batch;
uint32_t n_ubatch; uint32_t n_ubatch;
@ -17770,7 +17770,7 @@ static float is_graph_loaded(struct ggml_cgraph * cgraph) {
return float(n_loaded) / float(n_total) * 100.0f; return float(n_loaded) / float(n_total) * 100.0f;
} }
static void manage_graph_tensors(struct ggml_cgraph * cgraph, int advice, bool force = false) { static void manage_graph_tensors(struct ggml_cgraph * cgraph, int advice) {
long page_size = sysconf(_SC_PAGESIZE); long page_size = sysconf(_SC_PAGESIZE);
struct Segment { struct Segment {
@ -17826,8 +17826,8 @@ static void manage_graph_tensors(struct ggml_cgraph * cgraph, int advice, bool f
size_t prefetch_dense = 4; size_t prefetch_dense = 4;
size_t len = std::max(segment.end - segment.start, static_cast<size_t>(page_size)); 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 posix_madvise(reinterpret_cast<void *>(segment.start), len, advice); // hint to load into memory
// force to prefetch data // force to prefetch data, disabled by default
if (force && advice == POSIX_MADV_WILLNEED && false) { if (advice == POSIX_MADV_WILLNEED && false) {
volatile char * ptr = reinterpret_cast<volatile char *>(segment.start); volatile char * ptr = reinterpret_cast<volatile char *>(segment.start);
for (size_t off = 0; off < len; off += prefetch_dense * page_size) { for (size_t off = 0; off < len; off += prefetch_dense * page_size) {
for (size_t i = 0; i < prefetch_dense; i++) { for (size_t i = 0; i < prefetch_dense; i++) {
@ -18104,17 +18104,13 @@ static int llama_decode_internal(
} }
// overlap memory scheduling with other nodes' communication and computing // overlap memory scheduling with other nodes' communication and computing
{ if (cparams.prefetch && n_world > 1) {
timer(manage_graph_tensors); timer(manage_graph_tensors);
int next_gf_id = (i + 1) % gf.size(); int next_gf_id = (i + 1) % gf.size();
manage_graph_tensors(gf[next_gf_id], POSIX_MADV_WILLNEED, n_world > 1); manage_graph_tensors(gf[next_gf_id], POSIX_MADV_WILLNEED);
if (my_rank == 0 && (is_last_l || (next_gf_id == (int)gf.size() - 1))) { if (my_rank == 0 && (is_last_l || (next_gf_id == (int)gf.size() - 1))) {
manage_graph_tensors(gf[0], POSIX_MADV_WILLNEED, n_world > 1); manage_graph_tensors(gf[0], POSIX_MADV_WILLNEED);
}
if (cparams.unload && n_world > 1) {
manage_graph_tensors(sub_gf, POSIX_MADV_DONTNEED);
} }
} }
} }
@ -19837,7 +19833,7 @@ struct llama_context_params llama_context_default_params() {
/*.rank =*/ 0, /*.rank =*/ 0,
/*.n_layer_window =*/ {32}, /*.n_layer_window =*/ {32},
/*.n_gpu_layers =*/ 0, /*.n_gpu_layers =*/ 0,
/*.unload =*/ false, /*.prefetch =*/ false,
/*.keep_out_in_metal =*/ true, /*.keep_out_in_metal =*/ true,
/*.master_ip =*/ nullptr, /*.master_ip =*/ nullptr,
/*.next_node_ip =*/ nullptr, /*.next_node_ip =*/ nullptr,
@ -20265,7 +20261,7 @@ void * llama_context_setup_backend(
auto & cparams = ctx->cparams; auto & cparams = ctx->cparams;
std::copy(std::begin(params.n_layer_window), std::end(params.n_layer_window), cparams.n_layer_window); std::copy(std::begin(params.n_layer_window), std::end(params.n_layer_window), cparams.n_layer_window);
cparams.unload = params.unload; cparams.prefetch = params.prefetch;
cparams.n_seq_max = std::max(1u, params.n_seq_max); cparams.n_seq_max = std::max(1u, params.n_seq_max);
cparams.n_threads = params.n_threads; cparams.n_threads = params.n_threads;
cparams.n_threads_batch = params.n_threads_batch; cparams.n_threads_batch = params.n_threads_batch;