upstream-archive/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-07-23 11:10:55 -05:00
Code Issues Packages Projects Releases Wiki Activity

server : allow text-only slot save/restore with mtmd (#25076)

Browse Source
This commit is contained in:
Chipmunk
2026-07-16 21:26:44 +09:00
committed by GitHub
parent a55a8c5266
commit a8dc0e3269
3 changed files with 145 additions and 14 deletions
Download Patch File Download Diff File Expand all files Collapse all files

30
tools/server/server-context.cpp
View File

@@ -2011,10 +2011,13 @@ private:
queue_results.send(std::move(res));
}
// if multimodal is enabled, send an error and return false
bool check_no_mtmd(const int id_task) {
if (mctx) {
send_error(id_task, "This feature is not supported by multimodal", ERROR_TYPE_NOT_SUPPORTED);
// Gate slot save/restore/erase on slot content (does it hold media),
// not model capability: a multimodal model may hold a pure-text slot.
bool check_slot_no_media(const server_slot & slot, const int id_task) {
if (slot.prompt.tokens.has_media()) {
send_error(id_task,
"This operation is not supported while the slot holds image/audio tokens (a pure-text prefix is supported)",
ERROR_TYPE_NOT_SUPPORTED);
return false;
}
return true;
@@ -2502,16 +2505,15 @@ private:
} break;
case SERVER_TASK_TYPE_SLOT_SAVE:
{
if (!check_no_mtmd(task.id)) {
break;
}
const int id_slot = task.slot_action.id_slot;
server_slot * slot = get_slot_by_id(id_slot);
if (slot == nullptr) {
send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST);
break;
}
if (!check_slot_no_media(*slot, task.id)) {
break;
}
if (slot->is_processing()) {
// if requested slot is unavailable, we defer this task for processing later
SRV_DBG("requested slot is unavailable, defer task, id_task = %d\n", task.id);
@@ -2519,13 +2521,13 @@ private:
break;
}
const size_t token_count = slot->prompt.tokens.size();
const int64_t t_start = ggml_time_us();
std::string filename = task.slot_action.filename;
std::string filepath = task.slot_action.filepath;
const llama_tokens & tokens = slot->prompt.tokens.get_tokens();
const llama_tokens tokens = slot->prompt.tokens.get_text_tokens();
const size_t token_count = tokens.size();
const size_t nwrite = llama_state_seq_save_file(ctx_tgt, filepath.c_str(), slot->id, tokens.data(), token_count);
const int64_t t_end = ggml_time_us();
@@ -2543,7 +2545,6 @@ private:
} break;
case SERVER_TASK_TYPE_SLOT_RESTORE:
{
if (!check_no_mtmd(task.id)) break;
const int id_slot = task.slot_action.id_slot;
server_slot * slot = get_slot_by_id(id_slot);
if (slot == nullptr) {
@@ -2590,15 +2591,16 @@ private:
} break;
case SERVER_TASK_TYPE_SLOT_ERASE:
{
if (!check_no_mtmd(task.id)) {
break;
}
const int id_slot = task.slot_action.id_slot;
server_slot * slot = get_slot_by_id(id_slot);
if (slot == nullptr) {
send_error(task, "Invalid slot ID", ERROR_TYPE_INVALID_REQUEST);
break;
}
// Gate on slot content, consistent with save/restore.
if (!check_slot_no_media(*slot, task.id)) {
break;
}
if (slot->is_processing()) {
// if requested slot is unavailable, we defer this task for processing later
SRV_DBG("requested slot is unavailable, defer task, id_task = %d\n", task.id);