LLM inference in C/C++
C++ ggml
ggml : improve RPC tensor transfer logs (#30481)
Georgi Gerganovaf0579c78a07
diff --git a/ggml/src/ggml-rpc/CMakeLists.txt b/ggml/src/ggml-rpc/CMakeLists.txtindex e3d0c9b4c2..8976651238 100644--- a/ggml/src/ggml-rpc/CMakeLists.txt+++ b/ggml/src/ggml-rpc/CMakeLists.txt@@ -2,6 +2,7 @@ message(STATUS "Using RPC backend")ggml_add_backend_library(ggml-rpcggml-rpc.cpp+ log.cpptransport.cpp)diff --git a/ggml/src/ggml-rpc/ggml-rpc.cpp b/ggml/src/ggml-rpc/ggml-rpc.cppindex 45fa8a42d2..49f11073e9 100644--- a/ggml/src/ggml-rpc/ggml-rpc.cpp+++ b/ggml/src/ggml-rpc/ggml-rpc.cpp@@ -104,6 +104,11 @@ static const char * rpc_cmd_name(enum rpc_cmd cmd) {case RPC_CMD_DEVICE_COUNT: return "DEVICE_COUNT";case RPC_CMD_GRAPH_RECOMPUTE: return "GRAPH_RECOMPUTE";case RPC_CMD_MEMSET_TENSOR: return "MEMSET_TENSOR";+ case RPC_CMD_SET_TENSOR_2D: return "SET_TENSOR_2D";+ case RPC_CMD_GET_TENSOR_2D: return "GET_TENSOR_2D";+ case RPC_CMD_COMM_INIT: return "COMM_INIT";+ case RPC_CMD_COMM_ALLREDUCE: return "COMM_ALLREDUCE";+ case RPC_CMD_COMM_FREE: return "COMM_FREE";case RPC_CMD_NONE: return "NONE";default: return "UNKNOWN";}@@ -668,8 +673,9 @@ void rpc_dispatcher::work() {break;}if (msg_ptr->cmd != RPC_CMD_NONE) {- LOG_DBG2("[%s] %s (in: %zu, out: %zu)\n", __func__, rpc_cmd_name(msg_ptr->cmd), msg_ptr->input_size, msg_ptr->output_size);- const bool trace = rpc_debug_level() >= 3;+ LOG_DBG2("[%s] %s (in: %s, out: %s)\n", __func__, rpc_cmd_name(msg_ptr->cmd),+ rpc_log_format_size(msg_ptr->input_size).c_str(), rpc_log_format_size(msg_ptr->output_size).c_str());+ const bool trace = rpc_log_debug_level() >= 3;const auto t0 = std::chrono::steady_clock::now();if (msg_ptr->output) {bool status = send_rpc_cmd(sock, msg_ptr->cmd, msg_ptr->input.get(), msg_ptr->input_size, msg_ptr->output, msg_ptr->output_size);@@ -798,7 +804,7 @@ static enum ggml_status ggml_backend_rpc_buffer_init_tensor(ggml_backend_buffer_// Due to bandwidth constraints, we only call the server init tensor functions if necessary.// In particular, only quantized tensors need paddingif (ggml_is_quantized(tensor->type) && (tensor->ne[0] % 512 != 0) && (tensor->view_src == nullptr)) {- LOG_DBG2("[%s] tensor: %s\n", __func__, tensor->name);+ LOG_DBG2("[%s] tensor: %s\n", __func__, rpc_log_format_tensor(tensor).c_str());auto request = std::make_shared<rpc_msg_init_tensor_req>();request->tensor = serialize_tensor(tensor);ctx->dispatcher->send(RPC_CMD_INIT_TENSOR, request, sizeof(*request));@@ -814,7 +820,8 @@ static void ggml_backend_rpc_buffer_memset_tensor(request->offset = offset;request->size = size;request->value = value;- LOG_DBG2("[%s] tensor: %s, offset: %zu, size: %zu, value: %u\n", __func__, tensor->name, offset, size, value);+ LOG_DBG2("[%s] tensor: %s, offset: %zu, size: %s, value: %u\n", __func__, rpc_log_format_tensor(tensor).c_str(), offset,+ rpc_log_format_size(size).c_str(), value);ctx->dispatcher->send(RPC_CMD_MEMSET_TENSOR, request, sizeof(*request));}@@ -842,7 +849,8 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggmggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;rpc_tensor rpc_tensor = serialize_tensor(tensor);uint8_t cache_flag = 0;- LOG_DBG2("[%s] tensor: %s, offset: %zu, size: %zu\n", __func__, tensor->name, offset, size);+ LOG_DBG2("[%s] tensor: %s, offset: %zu, size: %s\n", __func__, rpc_log_format_tensor(tensor).c_str(), offset,+ rpc_log_format_size(size).c_str());if (rpc_use_hash_cache(tensor, size)) {auto request = std::make_shared<rpc_msg_set_tensor_hash_req>();request->tensor = rpc_tensor;@@ -850,7 +858,8 @@ static void ggml_backend_rpc_buffer_set_tensor(ggml_backend_buffer_t buffer, ggmrequest->hash = fnv_hash((const uint8_t*)data, size);rpc_msg_set_tensor_hash_rsp response;ctx->dispatcher->send(RPC_CMD_SET_TENSOR_HASH, request, sizeof(*request), &response, sizeof(response));- LOG_DBG2("[%s] tensor: %s, hash: 0x%" PRIx64 ", cache: %s\n", __func__, tensor->name, request->hash, response.result ? "hit" : "miss");+ LOG_DBG2("[%s] tensor: %s, hash: 0x%" PRIx64 ", cache: %s\n", __func__, rpc_log_format_tensor(tensor).c_str(),+ request->hash, response.result ? "hit" : "miss");if (response.result) {// the server has the same data, no need to send itreturn;@@ -867,6 +876,8 @@ static void ggml_backend_rpc_buffer_set_tensor_2d(ggml_backend_buffer_t buffer,size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data) {ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;rpc_tensor rpc_tensor = serialize_tensor(tensor);+ LOG_DBG2("[%s] tensor: %s, offset: %zu, size: %s, n_copies: %zu, stride: %zu\n", __func__,+ rpc_log_format_tensor(tensor).c_str(), offset, rpc_log_format_size(size).c_str(), n_copies, stride_tensor);// input serialization format: | rpc_tensor | offset (8 bytes) | size (8 bytes) | n_copies (8 bytes) | stride (8 bytes) | data (size * n_copies bytes) |size_t input_size = sizeof(rpc_tensor) + 4*sizeof(uint64_t) + size*n_copies;uint8_t * input = new uint8_t[input_size]();@@ -886,6 +897,8 @@ static void ggml_backend_rpc_buffer_set_tensor_2d(ggml_backend_buffer_t buffer,static void ggml_backend_rpc_buffer_get_tensor_2d(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data,size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data) {ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;+ LOG_DBG2("[%s] tensor: %s, offset: %zu, size: %s, n_copies: %zu, stride: %zu\n", __func__,+ rpc_log_format_tensor(tensor).c_str(), offset, rpc_log_format_size(size).c_str(), n_copies, stride_tensor);auto request = std::make_shared<rpc_msg_get_tensor_2d_req>();request->tensor = serialize_tensor(tensor);request->offset = offset;@@ -905,7 +918,8 @@ static void ggml_backend_rpc_buffer_get_tensor_2d(ggml_backend_buffer_t buffer,static void ggml_backend_rpc_buffer_get_tensor(ggml_backend_buffer_t buffer, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;- LOG_DBG2("[%s] tensor: %s, offset: %zu, size: %zu\n", __func__, tensor->name, offset, size);+ LOG_DBG2("[%s] tensor: %s, offset: %zu, size: %s\n", __func__, rpc_log_format_tensor(tensor).c_str(), offset,+ rpc_log_format_size(size).c_str());auto request = std::make_shared<rpc_msg_get_tensor_req>();request->tensor = serialize_tensor(tensor);request->offset = offset;@@ -925,7 +939,7 @@ static bool ggml_backend_rpc_buffer_cpy_tensor(ggml_backend_buffer_t buffer, conreturn false;}ggml_backend_rpc_buffer_context * ctx = (ggml_backend_rpc_buffer_context *)buffer->context;- LOG_DBG2("[%s] src: %s -> dst: %s\n", __func__, src->name, dst->name);+ LOG_DBG2("[%s] src: %s -> dst: %s\n", __func__, rpc_log_format_tensor(src).c_str(), rpc_log_format_tensor(dst).c_str());auto request = std::make_shared<rpc_msg_copy_tensor_req>();request->src = serialize_tensor(src);request->dst = serialize_tensor(dst);@@ -974,16 +988,17 @@ static ggml_backend_buffer_t ggml_backend_rpc_buffer_type_alloc_buffer(ggml_backauto dispatcher = get_dispatcher(buft_ctx->endpoint);dispatcher->send(RPC_CMD_ALLOC_BUFFER, request, sizeof(*request), &response, sizeof(response));if (response.remote_ptr != 0) {- LOG_DBG("[%s] endpoint: %s, device: %u, size: %" PRIu64 " -> remote_ptr: 0x%" PRIx64 ", remote_size: %" PRIu64 "\n",- __func__, buft_ctx->endpoint.c_str(), buft_ctx->device, request->size, response.remote_ptr, response.remote_size);+ LOG_DBG("[%s] endpoint: %s, device: %u, size: %s -> remote_ptr: 0x%" PRIx64 ", remote_size: %s\n",+ __func__, buft_ctx->endpoint.c_str(), buft_ctx->device, rpc_log_format_size(request->size).c_str(),+ response.remote_ptr, rpc_log_format_size(response.remote_size).c_str());ggml_backend_buffer_t buffer = ggml_backend_buffer_init(buft,ggml_backend_rpc_buffer_interface,new ggml_backend_rpc_buffer_context{dispatcher, nullptr, response.remote_ptr},response.remote_size);return buffer;} else {- LOG_WARN("[%s] endpoint: %s, device: %u, size: %" PRIu64 " -> failed\n",- __func__, buft_ctx->endpoint.c_str(), buft_ctx->device, request->size);+ LOG_WARN("[%s] endpoint: %s, device: %u, size: %s -> failed\n",+ __func__, buft_ctx->endpoint.c_str(), buft_ctx->device, rpc_log_format_size(request->size).c_str());return nullptr;}}@@ -993,7 +1008,7 @@ static size_t get_alignment(const std::shared_ptr<rpc_dispatcher> & dispatcher,request->device = device;rpc_msg_get_alignment_rsp response;dispatcher->send(RPC_CMD_GET_ALIGNMENT, request, sizeof(*request), &response, sizeof(response));- LOG_DBG2("[%s] device: %u, alignment: %" PRIu64 "\n", __func__, device, response.alignment);+ LOG_DBG2("[%s] device: %u, alignment: %s\n", __func__, device, rpc_log_format_size(response.alignment).c_str());return response.alignment;}@@ -1007,7 +1022,7 @@ static size_t get_max_size(const std::shared_ptr<rpc_dispatcher> & dispatcher, urequest->device = device;rpc_msg_get_max_size_rsp response;dispatcher->send(RPC_CMD_GET_MAX_SIZE, request, sizeof(*request), &response, sizeof(response));- LOG_DBG2("[%s] device: %u, max_size: %" PRIu64 "\n", __func__, device, response.max_size);+ LOG_DBG2("[%s] device: %u, max_size: %s\n", __func__, device, rpc_log_format_size(response.max_size).c_str());return response.max_size;}@@ -1073,7 +1088,8 @@ static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_tystd::lock_guard<std::mutex> lock(cache_mutex);auto it = cache.find(cache_hash);if (it != cache.end()) {- LOG_DBG3("[%s] cache hit: %s [%s] -> %zu\n", __func__, ggml_op_name(tensor->op), tensor->name, it->second);+ LOG_DBG3("[%s] cache hit: %s -> %s\n", __func__, rpc_log_format_tensor(tensor).c_str(),+ rpc_log_format_size(it->second).c_str());return std::max<size_t>(it->second, min_size);}}@@ -1087,7 +1103,7 @@ static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_tyrequest->srcs[i] = serialize_tensor(tensor->src[i]);}- LOG_DBG2("[%s] cache miss: %s [%s], querying server\n", __func__, ggml_op_name(tensor->op), tensor->name);+ LOG_DBG2("[%s] cache miss: %s, querying server\n", __func__, rpc_log_format_tensor(tensor).c_str());rpc_msg_get_alloc_size_rsp response;auto dispatcher = get_dispatcher(buft_ctx->endpoint);dispatcher->send(RPC_CMD_GET_ALLOC_SIZE, request, sizeof(*request), &response, sizeof(response));@@ -1097,7 +1113,8 @@ static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_tycache[cache_hash] = response.alloc_size;}- LOG_DBG2("[%s] %s [%s] -> alloc_size: %" PRIu64 "\n", __func__, ggml_op_name(tensor->op), tensor->name, response.alloc_size);+ LOG_DBG2("[%s] %s -> alloc_size: %s\n", __func__, rpc_log_format_tensor(tensor).c_str(),+ rpc_log_format_size(response.alloc_size).c_str());return std::max<size_t>(response.alloc_size, min_size);}@@ -1132,7 +1149,8 @@ static void ggml_backend_rpc_set_tensor_async(ggml_backend_t backend, ggml_tensoggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context;rpc_tensor rpc_tensor = serialize_tensor(tensor);uint8_t cache_flag = 0;- LOG_DBG2("[%s] tensor: %s, offset: %zu, size: %zu\n", __func__, tensor->name, offset, size);+ LOG_DBG2("[%s] tensor: %s, offset: %zu, size: %s\n", __func__, rpc_log_format_tensor(tensor).c_str(), offset,+ rpc_log_format_size(size).c_str());if (rpc_use_hash_cache(tensor, size)) {auto request = std::make_shared<rpc_msg_set_tensor_hash_req>();request->tensor = rpc_tensor;@@ -1141,7 +1159,8 @@ static void ggml_backend_rpc_set_tensor_async(ggml_backend_t backend, ggml_tensorpc_msg_set_tensor_hash_rsp response;// TODO: make this asyncctx->dispatcher->send(RPC_CMD_SET_TENSOR_HASH, request, sizeof(*request), &response, sizeof(response));- LOG_DBG2("[%s] tensor: %s, hash: 0x%" PRIx64 ", cache: %s\n", __func__, tensor->name, request->hash, response.result ? "hit" : "miss");+ LOG_DBG2("[%s] tensor: %s, hash: 0x%" PRIx64 ", cache: %s\n", __func__, rpc_log_format_tensor(tensor).c_str(),+ request->hash, response.result ? "hit" : "miss");if (response.result) {// the server has the same data, no need to send itreturn;@@ -1156,6 +1175,8 @@ static void ggml_backend_rpc_set_tensor_async(ggml_backend_t backend, ggml_tensostatic void ggml_backend_rpc_get_tensor_async(ggml_backend_t backend, const ggml_tensor * tensor, void * data, size_t offset, size_t size) {ggml_backend_rpc_context * ctx = (ggml_backend_rpc_context *)backend->context;+ LOG_DBG2("[%s] tensor: %s, offset: %zu, size: %s\n", __func__, rpc_log_format_tensor(tensor).c_str(), offset,+ rpc_log_format_size(size).c_str());auto request = std::make_shared<rpc_msg_get_tensor_req>();request->tensor = serialize_tensor(tensor);request->offset = offset;@@ -1240,7 +1261,7 @@ void rpc_dispatcher::graph_compute(uint32_t device, const ggml_cgraph * cgraph)}size_t input_size = 0;uint8_t * input = serialize_graph(device, cgraph, this, &input_size);- LOG_DBG2("[%s] serialized graph: %zu bytes\n", __func__, input_size);+ LOG_DBG2("[%s] serialized graph: %s\n", __func__, rpc_log_format_size(input_size).c_str());std::shared_ptr<uint8_t> input_ptr(input, std::default_delete<uint8_t[]>());send_async(RPC_CMD_GRAPH_COMPUTE, input_ptr, input_size);}@@ -1321,7 +1342,8 @@ ggml_backend_buffer_type_t ggml_backend_rpc_buffer_type(const char * endpoint, u/* .context = */ buft_ctx};buft_map[buft_name] = buft;- LOG_DBG("[%s] created %s (alignment: %zu, max_size: %zu)\n", __func__, buft_name.c_str(), alignment, max_size);+ LOG_DBG("[%s] created %s (alignment: %s, max_size: %s)\n", __func__, buft_name.c_str(),+ rpc_log_format_size(alignment).c_str(), rpc_log_format_size(max_size).c_str());return buft;}@@ -1355,7 +1377,8 @@ void ggml_backend_rpc_get_device_memory(const char * endpoint, uint32_t device,request->device = device;rpc_msg_get_device_memory_rsp response;dispatcher->send(RPC_CMD_GET_DEVICE_MEMORY, request, sizeof(*request), &response, sizeof(response));- LOG_DBG2("[%s] device: %u, free: %" PRIu64 ", total: %" PRIu64 "\n", __func__, device, response.free_mem, response.total_mem);+ LOG_DBG2("[%s] device: %u, free: %s, total: %s\n", __func__, device,+ rpc_log_format_size(response.free_mem).c_str(), rpc_log_format_size(response.total_mem).c_str());*free = response.free_mem;*total = response.total_mem;}@@ -1462,7 +1485,8 @@ bool rpc_server::get_alloc_size(const rpc_msg_get_alloc_size_req & request, rpc_}}- LOG_DBG2("[%s] device: %u, buffer: %p, data: %p\n", __func__, dev_id, (void*)tensor->buffer, tensor->data);+ LOG_DBG2("[%s] device: %u, tensor: %s, buffer: %p, data: %p\n", __func__, dev_id, rpc_log_format_tensor(tensor).c_str(),+ (void*)tensor->buffer, tensor->data);if (tensor->buffer == nullptr) {//No buffer allocated.buft = ggml_backend_get_default_buffer_type(backends[dev_id]);@@ -1487,11 +1511,12 @@ bool rpc_server::alloc_buffer(const rpc_msg_alloc_buffer_req & request, rpc_msg_if (buffer != nullptr) {response.remote_ptr = reinterpret_cast<uint64_t>(buffer);response.remote_size = buffer->size;- LOG_DBG("[%s] device: %u, size: %" PRIu64 " -> remote_ptr: 0x%" PRIx64 ", remote_size: %" PRIu64 "\n",- __func__, dev_id, request.size, response.remote_ptr, response.remote_size);+ LOG_DBG("[%s] device: %u, size: %s -> remote_ptr: 0x%" PRIx64 ", remote_size: %s\n",+ __func__, dev_id, rpc_log_format_size(request.size).c_str(), response.remote_ptr,+ rpc_log_format_size(response.remote_size).c_str());buffers.insert(buffer);} else {- LOG_WARN("[%s] device: %u, size: %" PRIu64 " -> failed\n", __func__, dev_id, request.size);+ LOG_WARN("[%s] device: %u, size: %s -> failed\n", __func__, dev_id, rpc_log_format_size(request.size).c_str());}return true;}@@ -1503,7 +1528,7 @@ bool rpc_server::get_alignment(const rpc_msg_get_alignment_req & request, rpc_ms}ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backends[dev_id]);size_t alignment = ggml_backend_buft_get_alignment(buft);- LOG_DBG2("[%s] device: %u, alignment: %zu\n", __func__, dev_id, alignment);+ LOG_DBG2("[%s] device: %u, alignment: %s\n", __func__, dev_id, rpc_log_format_size(alignment).c_str());response.alignment = alignment;return true;}@@ -1515,7 +1540,7 @@ bool rpc_server::get_max_size(const rpc_msg_get_max_size_req & request, rpc_msg_}ggml_backend_buffer_type_t buft = ggml_backend_get_default_buffer_type(backends[dev_id]);size_t max_size = ggml_backend_buft_get_max_size(buft);- LOG_DBG2("[%s] device: %u, max_size: %zu\n", __func__, dev_id, max_size);+ LOG_DBG2("[%s] device: %u, max_size: %s\n", __func__, dev_id, rpc_log_format_size(max_size).c_str());response.max_size = max_size;return true;}@@ -1606,8 +1631,9 @@ bool rpc_server::memset_tensor(const rpc_msg_memset_tensor_req & request) {return false;}- LOG_DBG2("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 ", value: %u\n",- __func__, (void *) tensor->buffer, tensor->data, request.offset, request.size, request.value);+ LOG_DBG2("[%s] tensor: %s, buffer: %p, data: %p, offset: %" PRIu64 ", size: %s, value: %u\n",+ __func__, rpc_log_format_tensor(tensor).c_str(), (void *) tensor->buffer, tensor->data, request.offset,+ rpc_log_format_size(request.size).c_str(), request.value);ggml_backend_tensor_memset(tensor, request.value, request.offset, request.size);return true;}@@ -1696,7 +1722,9 @@ bool rpc_server::set_tensor(const std::vector<uint8_t> & input) {LOG_ERROR("[%s] error deserializing tensor\n", __func__);return false;}- LOG_DBG2("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu\n", __func__, (void*)tensor->buffer, tensor->data, offset, size);+ LOG_DBG2("[%s] tensor: %s, buffer: %p, data: %p, offset: %" PRIu64 ", size: %s\n",+ __func__, rpc_log_format_tensor(tensor).c_str(), (void*)tensor->buffer, tensor->data, offset,+ rpc_log_format_size(size).c_str());// sanitize tensor->data{@@ -1757,8 +1785,9 @@ bool rpc_server::set_tensor_2d(const std::vector<uint8_t> & input) {LOG_ERROR("[%s] error deserializing tensor\n", __func__);return false;}- LOG_DBG2("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 ", n_copies: %" PRIu64 ", stride: %" PRIu64 "\n",- __func__, (void*)tensor->buffer, tensor->data, offset, size, n_copies, stride);+ LOG_DBG2("[%s] tensor: %s, buffer: %p, data: %p, offset: %" PRIu64 ", size: %s, n_copies: %" PRIu64 ", stride: %" PRIu64 "\n",+ __func__, rpc_log_format_tensor(tensor).c_str(), (void*)tensor->buffer, tensor->data, offset,+ rpc_log_format_size(size).c_str(), n_copies, stride);// sanitize tensor->data{@@ -1809,14 +1838,6 @@ bool rpc_server::get_cached_file(uint64_t hash, std::vector<uint8_t> & data) {bool rpc_server::set_tensor_hash(const rpc_msg_set_tensor_hash_req & request, rpc_msg_set_tensor_hash_rsp & response){sync_all_backends();- std::vector<uint8_t> cached_file;- if (!get_cached_file(request.hash, cached_file)) {- LOG_DBG2("[%s] hash: 0x%" PRIx64 ", cache miss\n", __func__, request.hash);- response.result = 0;- return true;- }- size_t size = cached_file.size();- LOG_DBG2("[%s] hash: 0x%" PRIx64 ", cache hit (%zu bytes)\n", __func__, request.hash, size);struct ggml_init_params params {/*.mem_size =*/ ggml_tensor_overhead(),/*.mem_buffer =*/ NULL,@@ -1830,8 +1851,18 @@ bool rpc_server::set_tensor_hash(const rpc_msg_set_tensor_hash_req & request, rpLOG_ERROR("[%s] error deserializing tensor\n", __func__);return false;}- LOG_DBG2("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %zu, hash: %" PRIx64 "\n",- __func__, (void*)tensor->buffer, tensor->data, request.offset, size, request.hash);++ std::vector<uint8_t> cached_file;+ if (!get_cached_file(request.hash, cached_file)) {+ LOG_DBG2("[%s] tensor: %s, offset: %" PRIu64 ", hash: 0x%" PRIx64 ", cache miss\n",+ __func__, rpc_log_format_tensor(tensor).c_str(), request.offset, request.hash);+ response.result = 0;+ return true;+ }+ size_t size = cached_file.size();+ LOG_DBG2("[%s] tensor: %s, buffer: %p, data: %p, offset: %" PRIu64 ", size: %s, hash: 0x%" PRIx64 ", cache hit\n",+ __func__, rpc_log_format_tensor(tensor).c_str(), (void*)tensor->buffer, tensor->data, request.offset,+ rpc_log_format_size(size).c_str(), request.hash);// sanitize tensor->data{@@ -1866,7 +1897,8 @@ bool rpc_server::init_tensor(const rpc_msg_init_tensor_req & request) {LOG_ERROR("Null tensor pointer passed to server init_tensor function.\n");return false;}- LOG_DBG2("[%s] buffer: %p, data: %p\n", __func__, (void*)tensor->buffer, tensor->data);+ LOG_DBG2("[%s] tensor: %s, buffer: %p, data: %p\n", __func__, rpc_log_format_tensor(tensor).c_str(),+ (void*)tensor->buffer, tensor->data);// Call the backend's buffer_init_tensor functionggml_backend_buffer_t buffer = tensor->buffer;if (buffer && buffer->iface.init_tensor) {@@ -1902,7 +1934,9 @@ bool rpc_server::get_tensor(const rpc_msg_get_tensor_req & request, std::vector<LOG_ERROR("[%s] error deserializing tensor\n", __func__);return false;}- LOG_DBG2("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 "\n", __func__, (void*)tensor->buffer, tensor->data, request.offset, request.size);+ LOG_DBG2("[%s] tensor: %s, buffer: %p, data: %p, offset: %" PRIu64 ", size: %s\n", __func__,+ rpc_log_format_tensor(tensor).c_str(), (void*)tensor->buffer, tensor->data, request.offset,+ rpc_log_format_size(request.size).c_str());// sanitize tensor->data{@@ -1938,8 +1972,9 @@ bool rpc_server::get_tensor_2d(const rpc_msg_get_tensor_2d_req & request, std::vLOG_ERROR("[%s] error deserializing tensor\n", __func__);return false;}- LOG_DBG2("[%s] buffer: %p, data: %p, offset: %" PRIu64 ", size: %" PRIu64 ", n_copies: %" PRIu64 ", stride: %" PRIu64 "\n",- __func__, (void*)tensor->buffer, tensor->data, request.offset, request.size, request.n_copies, request.stride);+ LOG_DBG2("[%s] tensor: %s, buffer: %p, data: %p, offset: %" PRIu64 ", size: %s, n_copies: %" PRIu64 ", stride: %" PRIu64 "\n",+ __func__, rpc_log_format_tensor(tensor).c_str(), (void*)tensor->buffer, tensor->data, request.offset,+ rpc_log_format_size(request.size).c_str(), request.n_copies, request.stride);// sanitize tensor->data{@@ -2006,8 +2041,7 @@ bool rpc_server::copy_tensor(const rpc_msg_copy_tensor_req & request, rpc_msg_coreturn false;}- LOG_DBG2("[%s] src->buffer: %p, dst->buffer: %p\n",- __func__, (void*) src->buffer, (void*) dst->buffer);+ LOG_DBG2("[%s] src: %s -> dst: %s\n", __func__, rpc_log_format_tensor(src).c_str(), rpc_log_format_tensor(dst).c_str());response.result = ggml_backend_buffer_copy_tensor(src, dst);return true;@@ -2147,7 +2181,7 @@ bool rpc_server::graph_compute(const std::vector<uint8_t> & input) {return false;}if (graph->nodes[i] != nullptr) {- LOG_DBG3("[%s] node %u: id=%" PRId64 ", op=%s, name=%s\n", __func__, i, id, ggml_op_name(graph->nodes[i]->op), graph->nodes[i]->name);+ LOG_DBG3("[%s] node %u: id=%" PRId64 ", %s\n", __func__, i, id, rpc_log_format_tensor(graph->nodes[i]).c_str());const size_t hash_pos = ggml_hash_insert(&graph->visited_hash_set, graph->nodes[i]);graph->use_counts[hash_pos] = tensor_ptrs.at(id)->use_count;}@@ -2386,7 +2420,8 @@ bool rpc_server::get_device_memory(const rpc_msg_get_device_memory_req & requestggml_backend_dev_memory(dev, &free, &total);response.free_mem = free;response.total_mem = total;- LOG_DBG2("[%s] device: %u, free: %" PRIu64 ", total: %" PRIu64 "\n", __func__, dev_id, response.free_mem, response.total_mem);+ LOG_DBG2("[%s] device: %u, free: %s, total: %s\n", __func__, dev_id,+ rpc_log_format_size(response.free_mem).c_str(), rpc_log_format_size(response.total_mem).c_str());return true;}diff --git a/ggml/src/ggml-rpc/log.cpp b/ggml/src/ggml-rpc/log.cppnew file mode 100644index 0000000000..89f0ce7051--- /dev/null+++ b/ggml/src/ggml-rpc/log.cpp@@ -0,0 +1,34 @@+#include "log.h"++#include <cinttypes>+#include <cstdio>++std::string rpc_log_format_size(size_t size) {+ static const char * units[] = { "B", "KiB", "MiB", "GiB", "TiB", "PiB" };++ double value = (double) size;+ size_t unit = 0;+ while (value >= 1024.0 && unit + 1 < sizeof(units) / sizeof(units[0])) {+ value /= 1024.0;+ unit++;+ }++ char buf[32];+ if (unit == 0) {+ snprintf(buf, sizeof(buf), "%zu B", size);+ } else {+ snprintf(buf, sizeof(buf), "%.2f %s", value, units[unit]);+ }+ return std::string(buf);+}++std::string rpc_log_format_tensor(const ggml_tensor * tensor) {+ char shape[64];+ snprintf(shape, sizeof(shape), "[%" PRId64 ", %" PRId64 ", %" PRId64 ", %" PRId64 "]", tensor->ne[0], tensor->ne[1],+ tensor->ne[2], tensor->ne[3]);++ char buf[GGML_MAX_NAME + 192];+ snprintf(buf, sizeof(buf), "%s [%s, %s, %s]", ggml_get_name(tensor), ggml_op_name(tensor->op),+ ggml_type_name(tensor->type), shape);+ return std::string(buf);+}diff --git a/ggml/src/ggml-rpc/log.h b/ggml/src/ggml-rpc/log.hindex d1105c8689..96a383d77d 100644--- a/ggml/src/ggml-rpc/log.h+++ b/ggml/src/ggml-rpc/log.h@@ -1,21 +1,16 @@#pragma once-// Log infrastructure for the RPC backend: all RPC logging goes through this header.-//-// LOG_ERROR / LOG_WARN / LOG_INFO are unconditional severity logs.-//-// LOG_DBG* are debug logs gated by the GGML_RPC_DEBUG verbosity variable:-// unset / 0 - disabled (only warnings and errors are printed)-// 1 - high-level events: connections, handshake, buffer allocations, graph computes, cache evictions-// 2 - per-command trace: every RPC message sent/received, tensor transfers, cache hits/misses-// 3 - fine detail: per-command timings, graph node dumps, transport byte counts-// non-numeric values fall back to 1+// Log infrastructure for the RPC backend: LOG_* are unconditional, LOG_DBG* are gated by GGML_RPC_DEBUG#include "ggml-impl.h"+#include <cstddef>#include <cstdlib>+#include <string>-static inline int rpc_debug_level() {+// GGML_RPC_DEBUG verbosity: unset / 0 disabled, 1 connections and allocations, 2 per-command trace,+// 3 per-message detail. Non-numeric values fall back to 1+static inline int rpc_log_debug_level() {static const int level = [] {const char * env = std::getenv("GGML_RPC_DEBUG");if (env == nullptr) {@@ -39,10 +34,16 @@ static inline int rpc_debug_level() {return level;}+// Byte counts in binary units, e.g. 8208 -> "8.02 KiB". Sizes below 1 KiB stay in plain bytes+std::string rpc_log_format_size(size_t size);++// Tensor identity: "blk.0.attn_q.weight [NONE, f16, [4096, 4096, 1, 1]]"+std::string rpc_log_format_tensor(const ggml_tensor * tensor);+#define LOG_INFO(...) GGML_LOG_INFO(__VA_ARGS__)#define LOG_WARN(...) GGML_LOG_WARN(__VA_ARGS__)#define LOG_ERROR(...) GGML_LOG_ERROR(__VA_ARGS__)-#define LOG_DBG(...) do { if (rpc_debug_level() >= 1) GGML_LOG_DEBUG(__VA_ARGS__); } while (0)-#define LOG_DBG2(...) do { if (rpc_debug_level() >= 2) GGML_LOG_DEBUG(__VA_ARGS__); } while (0)-#define LOG_DBG3(...) do { if (rpc_debug_level() >= 3) GGML_LOG_DEBUG(__VA_ARGS__); } while (0)+#define LOG_DBG(...) do { if (rpc_log_debug_level() >= 1) GGML_LOG_DEBUG(__VA_ARGS__); } while (0)+#define LOG_DBG2(...) do { if (rpc_log_debug_level() >= 2) GGML_LOG_DEBUG(__VA_ARGS__); } while (0)+#define LOG_DBG3(...) do { if (rpc_log_debug_level() >= 3) GGML_LOG_DEBUG(__VA_ARGS__); } while (0)diff --git a/ggml/src/ggml-rpc/transport.cpp b/ggml/src/ggml-rpc/transport.cppindex 0d5343853d..22b38e00f5 100644--- a/ggml/src/ggml-rpc/transport.cpp+++ b/ggml/src/ggml-rpc/transport.cpp@@ -532,7 +532,7 @@ bool socket_t::impl::rdma_recv(void * data, size_t size) {#endif // GGML_RPC_RDMAbool socket_t::impl::send_data(const void * data, size_t size) {- LOG_DBG3("[%s] transport: %s, size: %zu\n", __func__, use_rdma ? "RDMA" : "TCP", size);+ LOG_DBG3("[%s] transport: %s, size: %s\n", __func__, use_rdma ? "RDMA" : "TCP", rpc_log_format_size(size).c_str());#ifdef GGML_RPC_RDMA_APPLEif (use_rdma) {return rdma->send(data, size);@@ -557,7 +557,7 @@ bool socket_t::impl::send_data(const void * data, size_t size) {}bool socket_t::impl::recv_data(void * data, size_t size) {- LOG_DBG3("[%s] transport: %s, size: %zu\n", __func__, use_rdma ? "RDMA" : "TCP", size);+ LOG_DBG3("[%s] transport: %s, size: %s\n", __func__, use_rdma ? "RDMA" : "TCP", rpc_log_format_size(size).c_str());#ifdef GGML_RPC_RDMA_APPLEif (use_rdma) {return rdma->recv(data, size);