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

feat: fix the profiling logic

Browse Source
This commit is contained in:
DandinPower
2025-12-17 00:33:11 +08:00
parent c490e38ae7
commit eb7675f923
1 changed files with 15 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files
src/llama.cpp
+15 -4
View File
@@ -18786,13 +18786,16 @@ static int llama_decode_internal(
ubatch.activate_output = (my_rank == 0 && is_out_embd);
GGML_ASSERT(!(ubatch.activate_input && ubatch.activate_output));
std::string start_h2d_time = get_iso8601_ms_timestamp();
llama_set_inputs(lctx, ubatch);
std::string end_h2d_time = get_iso8601_ms_timestamp();
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);
}
ggml_backend_sched_synchronize(lctx.sched[i]); // sync before end time
std::string end_compute_time = get_iso8601_ms_timestamp();
sub_gf_out = ggml_graph_node(sub_gf, -1);
@@ -18809,10 +18812,6 @@ static int llama_decode_internal(
snprintf(layer_desc, sizeof(layer_desc), "transformer_blocks");
log_layer = true;
}
if (log_layer && lctx.cparams.enable_comm_compute_log) {
LLAMA_LOG_INFO("[%d][%s][compute][start][%s][sbatch_tokens: %lu, ubatch_tokens: %u]\n", my_rank, start_compute_time.c_str(), layer_desc, lctx.sbatch.n_tokens, ubatch.n_tokens);
LLAMA_LOG_INFO("[%d][%s][compute][end][%s][sbatch_tokens: %lu, ubatch_tokens: %u]\n", my_rank, end_compute_time.c_str(), layer_desc, lctx.sbatch.n_tokens, ubatch.n_tokens);
}
is_output = strcmp(sub_gf_out->name, "result_output") == 0;
if (is_output) {
@@ -18840,8 +18839,20 @@ static int llama_decode_internal(
ggml_backend_t backend = ggml_backend_sched_get_tensor_backend(lctx.sched[i], sub_gf_out);
GGML_ASSERT(buf_size <= ggml_nbytes(sub_gf_out));
GGML_ASSERT(backend != nullptr);
std::string start_d2h_time = get_iso8601_ms_timestamp();
ggml_backend_tensor_get_async(backend, sub_gf_out, embd_buf, 0, buf_size);
ggml_backend_sched_synchronize(lctx.sched[i]);
std::string end_d2h_time = get_iso8601_ms_timestamp();
if (log_layer && lctx.cparams.enable_comm_compute_log) {
LLAMA_LOG_INFO("[%d][%s][h2d][start][%s]\n", my_rank, start_h2d_time.c_str(), layer_desc);
LLAMA_LOG_INFO("[%d][%s][h2d][end][%s]\n", my_rank, end_h2d_time.c_str(), layer_desc);
LLAMA_LOG_INFO("[%d][%s][compute][start][%s]\n", my_rank, start_compute_time.c_str(), layer_desc);
LLAMA_LOG_INFO("[%d][%s][compute][end][%s]\n", my_rank, end_compute_time.c_str(), layer_desc);
LLAMA_LOG_INFO("[%d][%s][d2h][start][%s]\n", my_rank, start_d2h_time.c_str(), layer_desc);
LLAMA_LOG_INFO("[%d][%s][d2h][end][%s]\n", my_rank, end_d2h_time.c_str(), layer_desc);
}
// send the result to the next node or the master
if (!(n_world == 1 || (my_rank == 0 && is_last_l))) {