Skip to content

rpc : fix alloc size logic #17116

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
wants to merge 2 commits into
base: master
Choose a base branch
from
+39 −10
Open

rpc : fix alloc size logic #17116

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
e118dc9
rpc : fix alloc size logic
ggerganov Nov 9, 2025
493d0bd
rpc : bump version
ggerganov Nov 10, 2025
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
3 changes: 1 addition & 2 deletions ggml/include/ggml-rpc.h
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,13 +1,12 @@
#pragma once

#include "ggml.h"
#include "ggml-backend.h"

#ifdef __cplusplus
extern "C" {
#endif

#define RPC_PROTO_MAJOR_VERSION 3
#define RPC_PROTO_MAJOR_VERSION 4
#define RPC_PROTO_MINOR_VERSION 0
#define RPC_PROTO_PATCH_VERSION 0
#define GGML_RPC_MAX_SERVERS 16
Expand Down
46 changes: 38 additions & 8 deletions ggml/src/ggml-rpc/ggml-rpc.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -127,6 +127,7 @@ struct rpc_msg_device_count_rsp {
struct rpc_msg_get_alloc_size_req {
uint32_t device;
rpc_tensor tensor;
rpc_tensor srcs[GGML_MAX_SRC];
};

struct rpc_msg_get_alloc_size_rsp {
Expand Down Expand Up @@ -545,6 +546,11 @@ static void * ggml_backend_rpc_buffer_get_base(ggml_backend_buffer_t buffer) {

static rpc_tensor serialize_tensor(const ggml_tensor * tensor) {
rpc_tensor result;
if (!tensor) {
memset(&result, 0, sizeof(result));
return result;
}

result.id = reinterpret_cast<uint64_t>(tensor);
result.type = tensor->type;
if (tensor->buffer) {
Expand Down Expand Up @@ -726,23 +732,41 @@ static size_t ggml_backend_rpc_get_max_size(ggml_backend_buffer_type_t buft) {
}

static size_t ggml_backend_rpc_buffer_type_get_alloc_size(ggml_backend_buffer_type_t buft, const ggml_tensor * tensor) {
// should we query the remote server for the actual size
bool rpc_get = false;

// See comments in init_tensor.
if (ggml_is_quantized(tensor->type) && (tensor->ne[0] % 512 != 0) && (tensor->view_src == nullptr)) {
rpc_get |= ggml_is_quantized(tensor->type) && (tensor->ne[0] % 512 != 0) && (tensor->view_src == nullptr);

// ops that require additional memory for fleeting data on certain backends
// ref: https://github.com/ggml-org/llama.cpp/pull/15966
rpc_get |= tensor->op == GGML_OP_FLASH_ATTN_EXT;
rpc_get |= tensor->op == GGML_OP_MUL_MAT_ID;

if (rpc_get) {
ggml_backend_rpc_buffer_type_context * buft_ctx = (ggml_backend_rpc_buffer_type_context *)buft->context;
auto sock = get_socket(buft_ctx->endpoint);

rpc_msg_get_alloc_size_req request;
request.device = buft_ctx->device;
request.tensor = serialize_tensor(tensor);
rpc_msg_get_alloc_size_req request = {
/*.device =*/ buft_ctx->device,
/*.tensor =*/ serialize_tensor(tensor),
/*.srcs =*/ {},
};

// .get_alloc_size could be a function of the tensor's srcs, so we must serialize them as well
for (int i = 0; i < GGML_MAX_SRC; i++) {
request.srcs[i] = serialize_tensor(tensor->src[i]);
}

// TODO: cache the alloc responses to avoid extra RPC calls?
rpc_msg_get_alloc_size_rsp response;
bool status = send_rpc_cmd(sock, RPC_CMD_GET_ALLOC_SIZE, &request, sizeof(request), &response, sizeof(response));
RPC_STATUS_ASSERT(status);

return response.alloc_size;
} else {
return ggml_nbytes(tensor);
}

return ggml_nbytes(tensor);
}

static ggml_backend_buffer_type_i ggml_backend_rpc_buffer_type_interface = {
Expand Down Expand Up @@ -969,20 +993,26 @@ bool rpc_server::get_alloc_size(const rpc_msg_get_alloc_size_req & request, rpc_
}
ggml_backend_buffer_type_t buft;
struct ggml_init_params params {
/*.mem_size =*/ ggml_tensor_overhead(),
/*.mem_size =*/ ggml_tensor_overhead()*(1 + GGML_MAX_SRC),
/*.mem_buffer =*/ NULL,
/*.no_alloc =*/ true,
};

ggml_context_ptr ctx_ptr { ggml_init(params) };
GGML_ASSERT(ctx_ptr != nullptr);
ggml_context * ctx = ctx_ptr.get();
ggml_tensor * tensor = deserialize_tensor(ctx, &request.tensor);

ggml_tensor * tensor = deserialize_tensor(ctx, &request.tensor);
if (tensor == nullptr) {
GGML_LOG_ERROR("Null tensor pointer passed to server get_alloc_size function.\n");
return false;
}
for (int i = 0; i < GGML_MAX_SRC; i++) {
if (request.srcs[i].id != 0) {
tensor->src[i] = deserialize_tensor(ctx, &request.srcs[i]);
}
}

LOG_DBG("[%s] device: %d, buffer: %p, data: %p\n", __func__, dev_id, (void*)tensor->buffer, tensor->data);
if (tensor->buffer == nullptr) {
//No buffer allocated.
Expand Down
Loading