nikola/llama_cpp
1
0
Fork 0
Code Issues Pull Requests Actions 4 Packages Projects Releases Wiki Activity

mtmd : add post-decode callback (#24645)

Browse Source 
Assisted-by: pi:llama.cpp/Qwen3.6-27B
This commit is contained in:
Georgi Gerganov
2026-06-15 16:02:05 +03:00
committed by GitHub
parent 9dbc6621ae
commit e3cab403bf
5 changed files with 50 additions and 121 deletions
Download Patch File Download Diff File Expand all files Collapse all files

31
tools/server/server-common.cpp
View File

@@ -539,37 +539,6 @@ bool server_tokens::validate(const struct llama_context * ctx) const {
return true;
}
int32_t server_tokens::process_chunk(
llama_context * ctx,
mtmd_context * mctx,
size_t idx,
llama_pos pos,
int32_t seq_id,
size_t & n_tokens_out) const {
const auto & chunk = find_chunk(idx);
const char * name = mtmd_input_chunk_get_type(chunk.get()) == MTMD_INPUT_CHUNK_TYPE_IMAGE
? "image" : "audio";
SRV_INF("processing %s...\n", name);
int32_t n_batch = llama_n_batch(ctx);
int64_t t0 = ggml_time_ms();
llama_pos new_n_past; // unused for now
int32_t result = mtmd_helper_eval_chunk_single(mctx, ctx,
chunk.get(),
pos,
seq_id,
n_batch,
true, // logits last
&new_n_past);
SRV_INF("%s processed in %" PRId64 " ms\n", name, ggml_time_ms() - t0);
if (result != 0) {
LOG_ERR("mtmd_helper_eval failed with status %d", result);
n_tokens_out = 0;
return result;
}
n_tokens_out = mtmd_input_chunk_get_n_tokens(chunk.get());
return 0;
}
server_tokens server_tokens::clone() const {
server_tokens res;
res.has_mtmd = has_mtmd;