Public/koboldcpp
1
0
Fork 0
mirror of https://github.com/LostRuins/koboldcpp.git synced 2026-09-04 03:51:17 +02:00
Code Issues Packages Projects Releases Wiki Activity

args: add --video-* CLI arguments (#24318)

Browse Source 
* args: add --video-* CLI arguments

* gen docs

* nits

* add mtmd_helper_init_opt
This commit is contained in:
Xuan-Son Nguyen
2026-08-27 12:11:12 +02:00
committed by GitHub
parent 915dc6d38c
commit f29551215b
12 changed files with 138 additions and 48 deletions
Download Patch File Download Diff File Expand all files Collapse all files
tools/server/server-context.cpp
+19 -12
View File
@@ -794,6 +794,8 @@ public:
llama_model * model_tgt = nullptr;
mtmd_context * mctx = nullptr;
// note: video_params.ffmpeg_bin_dir points into params_base, which outlives this struct
mtmd_helper_init_opt init_opt = mtmd_helper_init_opt_default();
const llama_vocab * vocab = nullptr;
server_queue queue_tasks;
@@ -1118,6 +1120,11 @@ private:
}
SRV_INF("loaded multimodal model, '%s'\n", mmproj_path.c_str());
init_opt.video_params.fps_target = params_base.video_fps;
init_opt.video_params.timestamp_interval_ms = params_base.video_timestamp_interval_ms;
init_opt.video_params.ffmpeg_bin_dir = params_base.video_ffmpeg_bin_dir.empty()
? nullptr : params_base.video_ffmpeg_bin_dir.c_str();
if (params_base.ctx_shift) {
params_base.ctx_shift = false;
SRV_WRN("%s\n", "ctx_shift is not supported by multimodal, it will be disabled");
@@ -2134,9 +2141,9 @@ private:
try {
auto & prompt = task.cli_prompt;
if (mctx != nullptr) {
task.tokens = process_mtmd_prompt(mctx, prompt, task.cli_files);
task.tokens = process_mtmd_prompt(mctx, prompt, task.cli_files, init_opt);
} else {
task.tokens = std::move(tokenize_input_prompts(vocab, mctx, prompt, true, true)[0]);
task.tokens = std::move(tokenize_input_prompts(vocab, mctx, prompt, true, true, init_opt)[0]);
}
task.cli_prompt.clear();
task.cli_files.clear();
@@ -4165,10 +4172,10 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
if (res_type != TASK_RESPONSE_TYPE_NONE && ctx_server.mctx != nullptr) {
// This is the case used by OAI compatible chat path with MTMD. TODO It can be moved to the path below.
inputs.push_back(process_mtmd_prompt(ctx_server.mctx, prompt.get<std::string>(), files));
inputs.push_back(process_mtmd_prompt(ctx_server.mctx, prompt.get<std::string>(), files, ctx_server.init_opt));
} else {
// Everything else, including multimodal completions.
inputs = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true);
inputs = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true, ctx_server.init_opt);
}
// tasks.reserve(inputs.size()); // TODO: this is inaccurate due to child tasks
@@ -4752,7 +4759,7 @@ void server_routes::init_routes() {
data["input_extra"] = input_extra; // default to empty array if it's not exist
std::string prompt = json_value(data, "prompt", std::string());
std::vector<server_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, false, true);
std::vector<server_tokens> tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, false, true, ctx_server.init_opt);
SRV_DBG("creating infill tasks, n_prompts = %d\n", (int) tokenized_prompts.size());
data["prompt"] = format_prompt_infill(
ctx_server.vocab,
@@ -4816,7 +4823,7 @@ void server_routes::init_routes() {
};
this->post_chat_completions_tok = [this](const server_http_req & req) {
return handle_count_tokens(ctx_server.vocab, ctx_server.mctx, req, TASK_RESPONSE_TYPE_OAI_CHAT);
return handle_count_tokens(ctx_server.vocab, ctx_server.mctx, ctx_server.init_opt, req, TASK_RESPONSE_TYPE_OAI_CHAT);
};
this->post_control = [this](const server_http_req & req) {
@@ -4875,7 +4882,7 @@ void server_routes::init_routes() {
};
this->post_responses_tok_oai = [this](const server_http_req & req) {
return handle_count_tokens(ctx_server.vocab, ctx_server.mctx, req, TASK_RESPONSE_TYPE_OAI_RESP);
return handle_count_tokens(ctx_server.vocab, ctx_server.mctx, ctx_server.init_opt, req, TASK_RESPONSE_TYPE_OAI_RESP);
};
this->post_transcriptions_oai = [this](const server_http_req & req) {
@@ -4925,7 +4932,7 @@ void server_routes::init_routes() {
};
this->post_anthropic_count_tokens = [this](const server_http_req & req) {
return handle_count_tokens(ctx_server.vocab, ctx_server.mctx, req, TASK_RESPONSE_TYPE_ANTHROPIC);
return handle_count_tokens(ctx_server.vocab, ctx_server.mctx, ctx_server.init_opt, req, TASK_RESPONSE_TYPE_ANTHROPIC);
};
// same with handle_chat_completions, but without inference part
@@ -5058,7 +5065,7 @@ void server_routes::init_routes() {
std::vector<server_task> tasks;
tasks.reserve(documents.size());
for (size_t i = 0; i < documents.size(); i++) {
auto tmp = format_prompt_rerank(ctx_server.model_tgt, ctx_server.vocab, ctx_server.mctx, query, documents[i]);
auto tmp = format_prompt_rerank(ctx_server.model_tgt, ctx_server.vocab, ctx_server.mctx, query, documents[i], ctx_server.init_opt);
server_task task = server_task(SERVER_TASK_TYPE_RERANK);
task.id = rd.get_new_id();
task.tokens = std::move(tmp);
@@ -5296,7 +5303,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_embeddings_impl(cons
}
}
auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true);
auto tokenized_prompts = tokenize_input_prompts(ctx_server.vocab, ctx_server.mctx, prompt, true, true, ctx_server.init_opt);
for (const auto & tokens : tokenized_prompts) {
// this check is necessary for models that do not add BOS token to the input
if (tokens.empty()) {
@@ -5357,7 +5364,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_embeddings_impl(cons
return res;
}
std::unique_ptr<server_res_generator> server_routes::handle_count_tokens(const llama_vocab * vocab, mtmd_context * mctx, const server_http_req & req, task_response_type res_type) {
std::unique_ptr<server_res_generator> server_routes::handle_count_tokens(const llama_vocab * vocab, mtmd_context * mctx, const mtmd_helper_init_opt & init_opt, const server_http_req & req, task_response_type res_type) {
auto res = create_response();
std::vector<raw_buffer> files;
json body = json::parse(req.body);
@@ -5395,7 +5402,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_count_tokens(const l
if (!prompt.is_string()) {
throw std::runtime_error("for mtmd, input prompt must be a string.");
}
n_tokens = process_mtmd_prompt(mctx, prompt.get<std::string>(), files, true).size();
n_tokens = process_mtmd_prompt(mctx, prompt.get<std::string>(), files, init_opt, true).size();
} else {
n_tokens = tokenize_mixed(vocab, prompt, true, true).size();
}