CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

Add LLaDA-7b-MoE diffusion model (#16003)

Browse Source
This commit is contained in:
Aman Gupta
2025-09-16 10:38:28 +08:00
committed by GitHub
parent 3d4053f77f
commit 6d758839ff
9 changed files with 315 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
examples/diffusion/diffusion-cli.cpp
View File

@@ -510,19 +510,27 @@ static void diffusion_generate(llama_context * ctx,
n_generated = params.max_length;
}
static std::string format_input_text(const std::string & prompt, bool use_chat_template, llama_model * model) {
static std::string format_input_text(const std::string & prompt, const std::string & system_prompt, bool use_chat_template, llama_model * model) {
if (!use_chat_template) {
return prompt;
}
auto chat_templates = common_chat_templates_init(model, "");
common_chat_templates_inputs inputs;
common_chat_msg user_msg;
user_msg.role = "user";
user_msg.content = prompt;
inputs.add_generation_prompt = true;
common_chat_msg system_msg;
if (!system_prompt.empty()) {
system_msg.role = "system";
system_msg.content = system_prompt;
inputs.messages.push_back(system_msg);
}
common_chat_msg user_msg;
user_msg.role = "user";
user_msg.content = prompt;
inputs.messages.push_back(user_msg);
inputs.add_generation_prompt = true;
auto result = common_chat_templates_apply(chat_templates.get(), inputs);
@@ -579,7 +587,8 @@ int main(int argc, char ** argv) {
llama_set_n_threads(ctx, params.cpuparams.n_threads, params.cpuparams_batch.n_threads);
const llama_vocab * vocab = llama_model_get_vocab(model);
std::string formatted_prompt = format_input_text(params.prompt, params.enable_chat_template, model);
std::string formatted_prompt = format_input_text(params.prompt, params.system_prompt, params.enable_chat_template, model);
std::vector<llama_token> input_tokens = common_tokenize(vocab,
formatted_prompt,
@@ -596,6 +605,7 @@ int main(int argc, char ** argv) {
}
llama_token mask_token_id = llama_vocab_mask(vocab);
GGML_ASSERT(mask_token_id != LLAMA_TOKEN_NULL);
bool visual_mode = params.diffusion.visual_mode;