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

add feature: keep the forwarder if its previous device cannot directly connect to its next device.

Browse Source 
feat: nodes attempt connections during topology rebuild while preserving forwarders
This commit is contained in:
Zonghang Li
2025-06-12 16:57:35 +04:00
committed by GitHub
parent e50b3aa473 d4618de991
commit ba29717613
8 changed files with 299 additions and 76 deletions
Download Patch File Download Diff File Expand all files Collapse all files
Makefile
+8 -1
View File
@@ -950,7 +950,8 @@ OBJ_LLAMA = \
src/llama-grammar.o \
src/llama-sampling.o \
src/unicode.o \
src/unicode-data.o
src/unicode-data.o \
src/network-utils.o \
OBJ_COMMON = \
common/profiler.o \
@@ -1139,6 +1140,11 @@ src/unicode-data.o: \
src/unicode-data.cpp \
src/unicode-data.h
$(CXX) $(CXXFLAGS) -c $< -o $@
src/network-utils.o: \
src/network-utils.cpp \
src/network-utils.h
$(CXX) $(CXXFLAGS) -c $< -o $@
src/llama.o: \
src/llama.cpp \
@@ -1147,6 +1153,7 @@ src/llama.o: \
src/llama-grammar.h \
src/llama-sampling.h \
src/unicode.h \
src/network-utils.h \
include/llama.h \
ggml/include/ggml-cuda.h \
ggml/include/ggml-metal.h \
common/common.cpp
+34 -6
View File
@@ -28,6 +28,7 @@
#include <unordered_set>
#include <vector>
#include <thread>
#include <atomic>
#if defined(__APPLE__) && defined(__MACH__)
#include <sys/types.h>
@@ -1717,6 +1718,8 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
}
// sychronize device profile to the master node
NodeType node_type;
char is_forwarder[32] = {0};
if (my_rank == 0) {
if (auto_schedule) {
std::vector<device_info> dev_info_set(n_world);
@@ -1733,7 +1736,7 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
return iparams;
}
llama_bcast_layer_setup(lctx, n_layer_window, n_gpu_layers);
llama_rebuild_topo(lctx, n_layer_window, dev_info_set.data());
llama_rebuild_topo(lctx, n_layer_window, dev_info_set.data(), &node_type, is_forwarder);
} else {
// use the user-defined n_layer_window
std::copy(std::begin(params.n_layer_window), std::end(params.n_layer_window), n_layer_window);
@@ -1743,14 +1746,14 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
if (auto_schedule){
llama_send_device_info(lctx, &dev_info);
llama_recv_layer_setup(lctx, n_layer_window, n_gpu_layers);
llama_rebuild_topo (lctx, n_layer_window, nullptr);
llama_rebuild_topo (lctx, n_layer_window, nullptr, &node_type, is_forwarder);
} else {
llama_recv_layer_setup(lctx, n_layer_window, n_gpu_layers);
}
}
// if this is a weak device, then exit
if (n_layer_window[my_rank] <= 0) {
if (node_type == NodeType::NODE_TYPE_EXIT) {
LOG_INF("No layer is assigned to me, exit.\n");
llama_free(lctx);
llama_free_model(model);
@@ -1759,10 +1762,11 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
// update my rank and n_world
uint32_t update_rank = 0, update_n_world = 1;
uint32_t worker_rank = 0, n_worker = 1;
std::vector<uint32_t> n_layer_window_temp = {n_layer_window[0]}, n_gpu_layers_temp = {n_gpu_layers[0]};
for (uint32_t i = 1; i < n_world; i++) {
if (n_layer_window[i] <= 0) {
if (n_layer_window[i] <= 0 && is_forwarder[i] == 0) {
continue;
}
if (i <= my_rank) {
@@ -1771,6 +1775,13 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
update_n_world++;
n_layer_window_temp.push_back(n_layer_window[i]);
n_gpu_layers_temp.push_back(n_gpu_layers[i]);
if (n_layer_window[i] > 0) {
if (i <= my_rank) {
worker_rank++;
}
n_worker++;
}
}
memset(n_layer_window, 0, n_world * sizeof(uint32_t));
@@ -1793,8 +1804,25 @@ struct llama_init_result llama_init_from_gpt_params(gpt_params & params) {
params.n_world = update_n_world;
n_world = update_n_world;
llama_update_context_with_rankworld(lctx, update_rank, update_n_world);
llama_update_context_with_rankworld(lctx, update_rank, update_n_world, worker_rank, n_worker);
if(node_type == NodeType::NODE_TYPE_FORWARDER){
//just forward
std::atomic<bool> should_exit{false};
auto t = std::thread([lctx, &should_exit]() {
while(!should_exit) {
llama_forward_messages(lctx);
}
});
char * stop_signal = nullptr;
llama_free_sockets(lctx, &stop_signal); // this will block until receive stop signal
should_exit = true;
t.join();
exit(0);
}
// update n_layer_window and n_gpu_layers
std::copy(std::begin(n_layer_window), std::end(n_layer_window), params.n_layer_window);
std::copy(std::begin(n_layer_window), std::end(n_layer_window), cparams.n_layer_window);
common/profiler.cpp
+28 -2
View File
@@ -2621,7 +2621,7 @@ size_t serialize(const struct device_info * dev_info, char ** buffer) {
return total_size;
}
void deserialize(const char * buffer, struct device_info * dev_info) {
size_t deserialize(const char * buffer, struct device_info * dev_info) {
const char * ptr = buffer;
// rank
@@ -2821,6 +2821,32 @@ void deserialize(const char * buffer, struct device_info * dev_info) {
ptr += sizeof(float);
memcpy(&dev_info->gpu_props.cuda_mem_cpy_delay, ptr, sizeof(float));
ptr += sizeof(float);
// no need to synchronize model flops and model params
}
return ptr - buffer;
}
void TopoRebuildHelperInfo::deserialize(const char *buffer) {
size_t buffer_size = ::deserialize(buffer, &dev_info);
if (buffer_size == 0) {
LOG_ERR("%s: failed to deserialize device info\n", __func__);
return;
}
memcpy(&is_forwarder, buffer + buffer_size, 1);
}
size_t TopoRebuildHelperInfo::serialize(char **buffer) const{
size_t buffer_size = ::serialize(&dev_info, buffer);
char* buffer_ = (char*)malloc(buffer_size+1);
if (buffer_ == NULL) {
LOG_ERR("%s: failed to allocate %zu bytes for device info serialization\n",
__func__, buffer_size);
return 0;
}
memcpy(buffer_, *buffer, buffer_size);
memcpy(buffer_ + buffer_size, &is_forwarder, 1);
free(*buffer);
*buffer = buffer_;
return buffer_size + 1;
}
common/profiler.h
+13 -1
View File
@@ -346,6 +346,18 @@ struct device_info {
model_bytes() {}
};
struct TopoRebuildHelperInfo{
struct device_info dev_info;
char is_forwarder;
TopoRebuildHelperInfo():
dev_info(),
is_forwarder(0){}
void deserialize(const char * buffer);
size_t serialize(char ** buffer) const;
};
enum profiler_backend_type {
PROFILER_BACKEND_TYPE_CPU = 0,
PROFILER_BACKEND_TYPE_METAL = 1,
@@ -389,6 +401,6 @@ int device_has_blas (void);
int device_has_sycl (void);
size_t serialize (const struct device_info * dev_info, char ** buffer);
void deserialize(const char * buffer, struct device_info * dev_info);
size_t deserialize(const char * buffer, struct device_info * dev_info);
#endif // PROFILER_H
include/llama.h
+15 -2
View File
@@ -448,6 +448,12 @@ extern "C" {
struct llama_model_params params);
LLAMA_API void llama_free_model(struct llama_model * model);
enum NodeType{
NODE_TYPE_WORKER,
NODE_TYPE_FORWARDER,
NODE_TYPE_EXIT,
};
LLAMA_API void llama_init_sockets (struct llama_context * ctx, uint32_t n_world, uint32_t my_rank);
LLAMA_API void llama_free_sockets (struct llama_context * ctx, char ** msg);
@@ -455,7 +461,12 @@ extern "C" {
LLAMA_API int llama_send_device_info (struct llama_context * ctx, struct device_info * dev_info);
LLAMA_API int llama_bcast_startup_args(struct llama_context * ctx, uint32_t rank, struct startup_args * args);
LLAMA_API int llama_bcast_layer_setup (struct llama_context * ctx, uint32_t * n_layer_window, uint32_t * n_gpu_layers);
LLAMA_API int llama_rebuild_topo (struct llama_context * ctx, uint32_t * n_layer_window, struct device_info * dev_info_set);
LLAMA_API int llama_rebuild_topo (struct llama_context * ctx,
uint32_t * n_layer_window,
struct device_info * desv_info_set,
NodeType* node_type,
char * is_forwarder);
LLAMA_API int llama_forward_messages (struct llama_context * ctx);
LLAMA_API int llama_recv_layer_setup (struct llama_context * ctx, uint32_t * n_layer_window, uint32_t * n_gpu_layers);
LLAMA_API int llm_load_tensors(
@@ -466,7 +477,9 @@ extern "C" {
LLAMA_API void llama_update_context_with_rankworld(
struct llama_context * ctx,
uint32_t rank,
uint32_t n_world);
uint32_t n_world,
uint32_t worker_rank,
uint32_t n_worker);
LLAMA_API struct llama_context * llama_new_context_with_model(
struct llama_model * model,
src/llama.cpp
+168 -64
View File
@@ -11,6 +11,7 @@
#include "ggml-backend.h"
#include "profiler.h"
#include "network-utils.h"
#ifdef GGML_USE_RPC
# include "ggml-rpc.h"
@@ -173,12 +174,12 @@ static void zeros(std::ofstream & file, size_t n) {
}
// zmq helpers
static std::vector<zmq::message_t> dev_infos_to_messages(const device_info* infos,
uint32_t n_world){
static std::vector<zmq::message_t> topohelper_to_messages(const TopoRebuildHelperInfo* infos,
uint32_t n_world){
std::vector<zmq::message_t> res;
for (uint32_t i = 0; i < n_world; ++i) {
char * buffer = nullptr;
size_t buffer_size = serialize(&infos[i], &buffer);
size_t buffer_size = infos[i].serialize(&buffer);
res.emplace_back(buffer, buffer_size);
free(buffer);
}
@@ -2596,6 +2597,9 @@ static_assert(std::is_trivially_copyable<llama_hparams>::value, "llama_hparams m
struct llama_cparams {
uint32_t n_world;
uint32_t rank;
NodeType node_type;
uint32_t n_worker;
uint32_t worker_rank;
uint32_t original_next_rank; // original rank of the next node
uint32_t n_layer_window[32];
bool prefetch;
@@ -18212,6 +18216,9 @@ static int llama_decode_internal(
const uint32_t n_world = cparams.n_world;
const uint32_t my_rank = cparams.rank;
const uint32_t n_worker = cparams.n_worker;
const uint32_t worker_rank = cparams.worker_rank;
lctx.is_encoding = false;
const uint32_t n_tokens_all = batch_all.n_tokens;
if (my_rank != 0) {
@@ -18267,7 +18274,7 @@ static int llama_decode_internal(
sync_meta meta;
meta.n_ctx = cparams.n_ctx;
bool is_last_dev = (my_rank == n_world - 1);
bool is_last_dev = (worker_rank == n_worker - 1);
if (my_rank != 0) {
if (llama_recv_meta(*lctx.recv_socket, &meta) == -1) {
@@ -20428,6 +20435,27 @@ static uint32_t map_rank_to_port(uint32_t rank, uint32_t data_port) {
return data_port + rank;
}
static std::string try_connect(llama_context *ctx, uint32_t rank, TopoRebuildHelperInfo* infos, uint32_t n_world, zmq::socket_t** socket){
auto prv_rank = (rank - 1 + n_world) % n_world;
std::string ip = infos[prv_rank].dev_info.next_ip;
auto port = map_rank_to_port(rank, ctx->data_port);
if(!isPortOpen(ip, port)){
*socket = nullptr;
return "";
}
std::string send_endp = "tcp://" + ip + ":" + std::to_string(port);
*socket = new zmq::socket_t(*ctx->sock_context, zmq::socket_type::push);
try {
(*socket)->connect(send_endp);
} catch (const zmq::error_t& e) {
delete *socket;
*socket = nullptr;
return "";
}
return ip;
}
void llama_init_sockets(struct llama_context * ctx, uint32_t n_world, uint32_t my_rank) {
if (n_world == 1) {
return;
@@ -20602,95 +20630,165 @@ int llama_bcast_layer_setup(struct llama_context * ctx, uint32_t * n_layer_windo
return 0;
}
int llama_rebuild_topo(llama_context * ctx, uint32_t * n_layer_window, device_info * dev_info_set) {
int llama_rebuild_topo(llama_context * ctx,
uint32_t * n_layer_window,
device_info * dev_info_set,
NodeType * node_type,
char * is_forwarder) {
uint32_t n_world = ctx->cparams.n_world;
uint32_t my_rank = ctx->cparams.rank;
device_info * dev_info_ptr = nullptr;
TopoRebuildHelperInfo* topo_helper = new TopoRebuildHelperInfo[n_world];
if (dev_info_set == nullptr) {
if (dev_info_set == nullptr){
// for rank!=0, recv all devices info
std::vector<zmq::message_t> msgs;
if (!zmq::recv_multipart(*ctx->recv_socket, std::back_inserter(msgs))) {
return -1;
}
dev_info_ptr = new device_info[n_world];
for (size_t i = 0; i < msgs.size(); i++) {
deserialize((const char *)msgs[i].data(), &dev_info_ptr[i]);
topo_helper[i].deserialize((char *)msgs[i].data());
}
GGML_ASSERT(msgs.size() == n_world);
} else {
dev_info_ptr = dev_info_set;
for (size_t i = 0; i < n_world; i++) {
topo_helper[i].dev_info = dev_info_set[i];
topo_helper[i].is_forwarder = 0;
}
}
GGML_ASSERT(ctx != nullptr && ctx->send_socket != nullptr);
// notify next rank
auto next_rank = (my_rank + 1) % n_world;
if (n_layer_window[next_rank] <= 0 && next_rank != 0) {
auto next_connect_rank = (my_rank + 1) % n_world;
zmq::socket_t* socket_to_close = nullptr;
bool is_not_exit = n_layer_window[my_rank] > 0 || topo_helper[my_rank].is_forwarder == 1;
if (is_not_exit){
// reconstruct socket to the next valid rank
auto current_rank = my_rank;
std::vector<uint32_t> nodes;
auto next_rank_ = next_rank;
while (next_rank_ != my_rank) {
nodes.push_back(next_rank_);
if (n_layer_window[next_rank_] > 0) {
break;
}
next_rank_ = (next_rank_ + 1) % n_world;
current_rank = (current_rank + 1) % n_world;
}
if (next_rank_ == my_rank) {
// only one node
ctx->next_node_ip = "";
socket_to_close = ctx->send_socket;
ctx->send_socket = nullptr;
} else {
// iterate node reverse
zmq::socket_t* socket = nullptr;
std::string ip;
for (int i = nodes.size() - 1; i > 0; --i) {
auto rank = nodes[i];
ip = try_connect(ctx, rank, topo_helper, n_world, &socket);
if (!ip.empty()) {
next_connect_rank = rank;
break;
}
}
topo_helper[next_connect_rank].is_forwarder = 1;
if (next_connect_rank != next_rank) {
// reset socket
GGML_ASSERT(socket != nullptr);
GGML_ASSERT(!ip.empty());
socket_to_close = ctx->send_socket;
ctx->send_socket = socket;
ctx->next_node_ip = ip;
ctx->cparams.original_next_rank = next_connect_rank;
}
}
}else if (n_layer_window[next_rank] <= 0 && topo_helper[next_rank].is_forwarder == 0) {
socket_to_close = ctx->send_socket;
}
// notify next exiting node
if (socket_to_close != nullptr) {
GGML_ASSERT(n_layer_window[next_rank] <= 0 && topo_helper[next_rank].is_forwarder == 0);
try {
auto msgs = dev_infos_to_messages(dev_info_ptr, n_world);
ctx->send_socket->set(zmq::sockopt::linger, 3500);
zmq::send_multipart(*ctx->send_socket, msgs);
auto msgs = topohelper_to_messages(topo_helper, n_world);
socket_to_close->set(zmq::sockopt::linger, 3500);
zmq::send_multipart(*socket_to_close, msgs);
} catch (const zmq::error_t& e) {
LLAMA_LOG_INFO("Failed to send data: %s\n", e.what());
if(!dev_info_set){
delete[] dev_info_ptr;
}
return -1;
}
}
zmq::socket_t * socket_to_close = nullptr;
if (n_layer_window[my_rank] > 0) {
// reconstruct socket to the next valid rank
std::string next_ip;
auto current_rank = my_rank;
while (next_rank != my_rank) {
if (n_layer_window[next_rank] > 0) {
next_ip = dev_info_ptr[current_rank].next_ip;
break;
}
next_rank = (next_rank + 1) % n_world;
current_rank = (current_rank + 1) % n_world;
}
if (!next_ip.empty()) {
if ((my_rank + 1) % n_world != next_rank) {
socket_to_close = ctx->send_socket;
ctx->send_socket = new zmq::socket_t(*ctx->sock_context, zmq::socket_type::push);
std::string send_endp = "tcp://" + next_ip + ":" + std::to_string(map_rank_to_port(next_rank, ctx->data_port));
ctx->send_socket->connect(send_endp);
ctx->next_node_ip = next_ip;
ctx->cparams.original_next_rank = next_rank;
}
if (next_rank != 0) {
try {
auto msgs = dev_infos_to_messages(dev_info_ptr, n_world);
zmq::send_multipart(*ctx->send_socket, msgs);
} catch (const zmq::error_t &e) {
LLAMA_LOG_INFO("Error binding/connecting recv socket to endpoint: %s", e.what());
if(!dev_info_set){
delete[] dev_info_ptr;
}
return -1;
}
}
} else {
// only one node
ctx->next_node_ip = "";
// notify next connect node
if(!ctx->next_node_ip.empty() && is_not_exit){
GGML_ASSERT(ctx->send_socket != nullptr);
try {
auto msgs = topohelper_to_messages(topo_helper, n_world);
zmq::send_multipart(*ctx->send_socket, msgs);
} catch (const zmq::error_t& e) {
LLAMA_LOG_INFO("Failed to send data: %s\n", e.what());
return -1;
}
}
if (!dev_info_set) {
delete[] dev_info_ptr;
if(n_layer_window[my_rank] > 0){
*node_type = NodeType::NODE_TYPE_WORKER;
}else if (topo_helper[my_rank].is_forwarder == 1){
*node_type = NodeType::NODE_TYPE_FORWARDER;
}else{
*node_type = NodeType::NODE_TYPE_EXIT;
}
if(ctx->send_socket != nullptr && *node_type!=NodeType::NODE_TYPE_EXIT){
// recv the whole view of all nodes
std::vector<zmq::message_t> msgs;
if (!zmq::recv_multipart(*ctx->recv_socket, std::back_inserter(msgs))) {
return -1;
}
GGML_ASSERT(msgs.size() == n_world);
for (size_t i = 0; i < msgs.size(); i++) {
topo_helper[i].deserialize((char *)msgs[i].data());
}
// broadcast the whole view
if(next_connect_rank!=0){
try {
zmq::send_multipart(*ctx->send_socket, msgs);
} catch (const zmq::error_t& e) {
LLAMA_LOG_INFO("Failed to send data: %s\n", e.what());
return -1;
}
}
}
for(size_t i = 0; i < n_world; i++) {
is_forwarder[i] = topo_helper[i].is_forwarder;
}
ctx->cparams.node_type = *node_type;
if(socket_to_close != nullptr){
if (socket_to_close != nullptr) {
socket_to_close->close();
delete socket_to_close;
}
delete [] topo_helper;
return 0;
}
int llama_forward_messages(llama_context *ctx) {
zmq::message_t message;
int more = true;
int timeout_ms = 10;
ctx->recv_socket->setsockopt(ZMQ_RCVTIMEO, &timeout_ms, sizeof(timeout_ms));
while (more) {
auto recv_result = ctx->recv_socket->recv(message, zmq::recv_flags::none);
if (!recv_result) {
return -1;
}
size_t more_size = sizeof(more);
ctx->recv_socket->getsockopt(ZMQ_RCVMORE, &more, &more_size);
ctx->send_socket->send(message,
more ? zmq::send_flags::sndmore : zmq::send_flags::none);
}
return 0;
}
@@ -20755,10 +20853,16 @@ void llama_free_sockets(struct llama_context * ctx, char ** msg) {
}
}
void llama_update_context_with_rankworld(struct llama_context * ctx, uint32_t rank, uint32_t n_world) {
void llama_update_context_with_rankworld(struct llama_context * ctx,
uint32_t rank,
uint32_t n_world,
uint32_t worker_rank,
uint32_t n_worker) {
if (ctx) {
ctx->cparams.rank = rank;
ctx->cparams.n_world = n_world;
ctx->cparams.worker_rank = worker_rank;
ctx->cparams.n_worker = n_worker;
}
}
src/network-utils.cpp
+26
View File
@@ -0,0 +1,26 @@
#include "network-utils.h"
#include <sys/socket.h>
#include <netinet/in.h>
#include <arpa/inet.h>
#include <unistd.h>
bool isPortOpen(const std::string& ip, uint32_t port, int timeout_sec) {
int sock = socket(AF_INET, SOCK_STREAM, 0);
if (sock < 0) return false;
struct timeval tv;
tv.tv_sec = timeout_sec;
tv.tv_usec = 0;
setsockopt(sock, SOL_SOCKET, SO_RCVTIMEO, &tv, sizeof(tv));
setsockopt(sock, SOL_SOCKET, SO_SNDTIMEO, &tv, sizeof(tv));
struct sockaddr_in server;
server.sin_addr.s_addr = inet_addr(ip.c_str());
server.sin_family = AF_INET;
server.sin_port = htons(port);
int res = connect(sock, (struct sockaddr*)&server, sizeof(server));
close(sock);
return res == 0;
}
src/network-utils.h
+7
View File
@@ -0,0 +1,7 @@
#pragma once
#include <string>
typedef unsigned int uint32_t;
bool isPortOpen(const std::string& ip, uint32_t port, int timeout_sec = 2);