vrr/prima.cpp
2
0
Fork 0
mirror of https://github.com/Lizonghang/prima.cpp.git synced 2025-09-10 15:54:34 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

feat(comm): Enhance robustness of communication

Browse source
This commit is contained in:
leeetao 2025-07-19 07:57:57 +00:00
parent 50a916f123
commit 663ad2896d
2 changed files with 45 additions and 7 deletions
Download patch file Download diff file Expand all files Collapse all files

50
examples/perplexity/perplexity.cpp
View file

@ -632,18 +632,39 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
const auto t_start = std::chrono::high_resolution_clock::now();
{
// synvhronize the KV cache clear signal across all ranks
// synchronize the KV cache clear signal across all ranks
if (n_world > 1) {
sync_meta clear_meta;
clear_meta.clear_kv_cache = true;
if (my_rank == 0) {
// Rank 0 sends the signal
llama_send_meta(ctx, &clear_meta);
} else {
if (llama_recv_meta(ctx, &clear_meta) == -1) {
LOG_ERR("Failed to recv clear_kv_cache signal on rank %d\n", my_rank);
// Non-rank 0 devices receive the signal with retry mechanism
int retry_count = 0;
const int max_retries = 5;
bool recv_success = false;
while (retry_count < max_retries && !recv_success) {
if (llama_recv_meta(ctx, &clear_meta) == 0) {
recv_success = true;
} else {
retry_count++;
LOG_WRN("Failed to recv clear_kv_cache signal on rank %d, retry %d/%d\n",
my_rank, retry_count, max_retries);
// Small delay before retry
std::this_thread::sleep_for(std::chrono::milliseconds(100));
}
}
if (!recv_success) {
LOG_ERR("Failed to recv clear_kv_cache signal on rank %d after %d retries\n",
my_rank, max_retries);
return {tokens, -1.0, {}, {}};
}
// Forward the signal to next rank if not the last device
if (!is_last_dev) {
llama_send_meta(ctx, &clear_meta);
}
@ -707,9 +728,26 @@ static results_perplexity perplexity(llama_context * ctx, const gpt_params & par
// Chain forwarding pattern: consistent with tokens_size communication
if (n_world > 1) {
if (my_rank != 0) {
// Non-rank 0 devices receive batch meta data
if (llama_recv_meta(ctx, &meta) == -1) {
LOG_ERR("Failed to recv batch meta on rank %d\n", my_rank);
// Non-rank 0 devices receive batch meta data with retry mechanism
int retry_count = 0;
const int max_retries = 3;
bool recv_success = false;
while (retry_count < max_retries && !recv_success) {
if (llama_recv_meta(ctx, &meta) == 0) {
recv_success = true;
} else {
retry_count++;
LOG_WRN("Failed to recv batch meta on rank %d, retry %d/%d\n",
my_rank, retry_count, max_retries);
// Small delay before retry
std::this_thread::sleep_for(std::chrono::milliseconds(50));
}
}
if (!recv_success) {
LOG_ERR("Failed to recv batch meta on rank %d after %d retries\n",
my_rank, max_retries);
return {tokens, -1.0, {}, {}};
}