electron-rare/prima-cpp
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

feat: add comm & compute log for further gantt chart analysis

Browse Source
This commit is contained in:
DandinPower
2025-07-05 18:09:11 +08:00
parent f1f7e37cdd
commit a399e49194
1 changed files with 37 additions and 1 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama.cpp
+37 -1
View File
@@ -92,6 +92,18 @@
#include <chrono>
#include <regex>
#include <inttypes.h>
#include <iomanip>
std::string get_iso8601_ms_timestamp() {
auto now = std::chrono::system_clock::now();
auto now_ms = std::chrono::duration_cast<std::chrono::milliseconds>(now.time_since_epoch()) % 1000;
auto now_c = std::chrono::system_clock::to_time_t(now);
std::tm tm_utc = *std::gmtime(&now_c);
std::ostringstream oss;
oss << std::put_time(&tm_utc, "%Y-%m-%dT%H:%M:%S");
oss << '.' << std::setw(3) << std::setfill('0') << now_ms.count() << 'Z';
return oss.str();
}
#if defined(_MSC_VER)
#pragma warning(disable: 4244 4267) // possible loss of data
@@ -18489,7 +18501,9 @@ static int llama_decode_internal(
// receive data from other nodes
if (n_world > 1 && !(my_rank == 0 && i == 0) && !(my_rank == 0 && is_last_l)) {
LLAMA_LOG_INFO("[%d][%s][comm][start][recv_tensors][n_tokens: %u, receive data from other nodes]\n", my_rank, get_iso8601_ms_timestamp().c_str(), n_tokens_all);
llama_recv_tensors(*lctx.recv_socket, &ubatch, is_out_embd);
LLAMA_LOG_INFO("[%d][%s][comm][end][recv_tensors][receive data from other nodes]\n", my_rank, get_iso8601_ms_timestamp().c_str());
}
// ensure ggml_backend_tensor_get_async of the previous subgraph has finished
@@ -18503,12 +18517,32 @@ static int llama_decode_internal(
llama_set_inputs(lctx, ubatch);
std::string start_compute_time = get_iso8601_ms_timestamp();
{ // compute graph
timer(llama_graph_compute);
llama_graph_compute(lctx, sub_gf, lctx.sched[i], n_threads, threadpool);
}
std::string end_compute_time = get_iso8601_ms_timestamp();
sub_gf_out = ggml_graph_node(sub_gf, -1);
bool log_layer = false;
char layer_desc[32];
if (strcmp(sub_gf_out->name, "inp_embd") == 0) {
snprintf(layer_desc, sizeof(layer_desc), "input_embedding");
log_layer = true;
} else if (strcmp(sub_gf_out->name, "result_output") == 0) {
snprintf(layer_desc, sizeof(layer_desc), "output_linear");
log_layer = true;
} else if (strncmp(sub_gf_out->name, "l_out", 5) == 0) {
snprintf(layer_desc, sizeof(layer_desc), "transformer_blocks");
log_layer = true;
}
if (log_layer) {
LLAMA_LOG_INFO("[%d][%s][compute][start][%s][N/A]\n", my_rank, start_compute_time.c_str(), layer_desc);
LLAMA_LOG_INFO("[%d][%s][compute][end][%s][N/A]\n", my_rank, end_compute_time.c_str(), layer_desc);
}
is_output = strcmp(sub_gf_out->name, "result_output") == 0;
if (is_output) {
break;
@@ -18540,10 +18574,12 @@ static int llama_decode_internal(
// send the result to the next node or the master
if (!(n_world == 1 || (my_rank == 0 && is_last_l))) {
LLAMA_LOG_INFO("[%d][%s][comm][start][send_tensors][send the result to the next node or the master]\n", my_rank, get_iso8601_ms_timestamp().c_str());
struct input_tensors tensors = {sub_gf_out, lctx.inp_pos};
const bool is_to_master = my_rank != 0 && is_last_l;
zmq::socket_t * s = is_to_master ? lctx.master_socket : lctx.send_socket;
llama_send_tensors(*s, &ubatch, &tensors);
LLAMA_LOG_INFO("[%d][%s][comm][end][send_tensors][send the result to the next node or the master]\n", my_rank, get_iso8601_ms_timestamp().c_str());
}
// overlap memory scheduling with other nodes' communication and computing