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

mtmd: add batching API (#24384)

Browse Source 
* mtmd: add batching API

* wip

* first working version (gemma4v)

* add arg

* nits

* wire up support_batch()

* fix 0.0 output embd

* fix audio

* nits

* refactor a bit

* nits

* fix non-batching case

* fix comment
This commit is contained in:
Xuan-Son Nguyen
2026-06-13 00:10:29 +02:00
committed by GitHub
parent f58bad4137
commit e37abd6b5f
14 changed files with 544 additions and 126 deletions
Download Patch File Download Diff File Expand all files Collapse all files

123
tools/server/server-context.cpp
View File

@@ -80,6 +80,8 @@ struct server_slot {
// multimodal
mtmd_context * mctx = nullptr;
mtmd::batch_ptr mbatch = nullptr;
std::array<llama_context *, 2> mtgt = {nullptr, nullptr}; // [0] for main context, [1] for optional draft context
// speculative decoding
common_speculative * spec;
@@ -239,6 +241,18 @@ struct server_slot {
// clear alora start
alora_invocation_start = -1;
// clear multimodal state
mbatch.reset();
mtgt[0] = ctx_tgt;
mtgt[1] = nullptr;
if (ctx_dft && llama_get_ctx_other(ctx_dft) != ctx_tgt) {
// TODO: in the future, figure out how to infuse target embeddings to the images
// for now, we re-decode the same chunk in both ctx_tgt and ctx_dft
// maybe we simply need to call `common_speculative_process()` ?
// [TAG_MTMD_DRAFT_PROCESSING]
mtgt[1] = ctx_dft;
}
}
void init_sampler() const {
@@ -578,6 +592,87 @@ struct server_slot {
other.prompt = prompt.clone();
other.init_sampler();
}
// returns 0 on success
// caller need to update prompt.tokens after a successful call to keep track of the processing progress
int process_mtmd_chunk(size_t idx, size_t & n_tokens_out) {
GGML_ASSERT(mctx);
const auto & input_tokens = task->tokens;
auto & chunk = input_tokens.find_chunk(idx);
int32_t res = 0;
auto try_decode = [&]() -> int32_t {
if (mbatch) {
float * embd = mtmd_batch_get_output_embd(mbatch.get(), chunk.get());
if (embd) {
for (auto * lctx : mtgt) {
if (lctx == nullptr) {
continue;
}
llama_pos new_n_past; // unused for now
res = mtmd_helper_decode_image_chunk(
mctx,
lctx,
chunk.get(),
embd,
prompt.tokens.pos_next(),
id,
llama_n_batch(lctx),
&new_n_past
);
if (res != 0) {
SLT_ERR(*this, "failed to decode mtmd chunk, idx = %zu, res = %d\n", idx, res);
return -1;
}
}
n_tokens_out = mtmd_input_chunk_get_n_tokens(chunk.get());
return 0; // success
}
}
return 1; // (non-error) need to create & encode batch
};
// if the batch is already exist, try searching & encode
res = try_decode();
if (res == 0) {
return 0;
} else if (res < 0) {
// fatal error
return res;
}
// otherwise, the batch is either uninitialized or is used up
// we need to create & encode a new batch
mbatch.reset(mtmd_batch_init(mctx));
res = mtmd_batch_add_chunk(mbatch.get(), chunk.get());
GGML_ASSERT(res == 0); // we should never have an empty batch
// try batching as much as possible
int n_added = 1;
size_t idx_cur = idx;
while (res == 0) {
auto [next_chunk, next_idx] = input_tokens.find_next_media_chunk(idx_cur);
if (next_chunk == nullptr) {
break;
}
res = mtmd_batch_add_chunk(mbatch.get(), next_chunk->get());
n_added += (res == 0 ? 1 : 0);
idx_cur = next_idx;
SLT_DBG(*this, "try adding media chunk idx = %zu to batch, res = %d\n", next_idx, res);
// if res != 0, batch is full or chunk is not compatible -> this loop breaks
}
// TODO @ngxson : move this log line to debug when it become more stable
SLT_INF(*this, "encoding mtmd batch from idx = %zu, n_chunks = %d\n", idx, n_added);
res = mtmd_batch_encode(mbatch.get());
if (res != 0) {
SLT_ERR(*this, "failed to encode mtmd batch for chunk idx = %zu, res = %d\n", idx, res);
return -1;
}
return try_decode();
}
};
@@ -781,6 +876,7 @@ private:
mparams.warmup = params_base.warmup;
mparams.image_min_tokens = params_base.image_min_tokens;
mparams.image_max_tokens = params_base.image_max_tokens;
mparams.batch_max_tokens = params_base.mtmd_batch_max_tokens;
mparams.media_marker = get_media_marker();
}
@@ -2928,7 +3024,7 @@ private:
send_partial_response(slot, {}, false, true);
}
}
}
} // end of SLOT_STATE_STARTED
if (!slot.can_split()) {
// cannot fit the prompt in the current batch - will try next iter
@@ -2983,10 +3079,18 @@ private:
bool has_mtmd = false;
// check if we should process the image
while (slot.prompt.n_tokens() < slot.task->n_tokens() && input_tokens[slot.prompt.n_tokens()] == LLAMA_TOKEN_NULL) {
while (true) {
auto cur_token_idx = slot.prompt.n_tokens();
if (
cur_token_idx >= slot.task->n_tokens() ||
input_tokens[cur_token_idx] != LLAMA_TOKEN_NULL // encountered a text token
) {
break;
}
// process the image
size_t n_tokens_out = 0;
int32_t res = input_tokens.process_chunk(ctx_tgt, mctx, slot.prompt.n_tokens(), slot.prompt.tokens.pos_next(), slot.id, n_tokens_out);
int32_t res = slot.process_mtmd_chunk(cur_token_idx, n_tokens_out);
if (res != 0) {
SLT_ERR(slot, "failed to process image, res = %d\n", res);
send_error(slot, "failed to process image", ERROR_TYPE_SERVER);
@@ -2994,22 +3098,11 @@ private:
continue;
}
if (ctx_dft && llama_get_ctx_other(ctx_dft.get()) != ctx_tgt) {
// TODO: in the future, figure out how to infuse target embeddings to the images
// for now, we skip this for simplicity
// maybe we simply need to call `common_speculative_process()` on the mtmd batches in the `process_chunk` above?
// [TAG_MTMD_DRAFT_PROCESSING]
res = input_tokens.process_chunk(ctx_dft.get(), mctx, slot.prompt.n_tokens(), slot.prompt.tokens.pos_next(), slot.id, n_tokens_out);
if (res != 0) {
GGML_ABORT("failed to process multi-modal data on draft context\n");
}
}
slot.n_prompt_tokens_processed += n_tokens_out;
// add the image chunk to cache
{
const auto & chunk = input_tokens.find_chunk(slot.prompt.n_tokens());
const auto & chunk = input_tokens.find_chunk(cur_token_idx);
slot.prompt.tokens.push_back(chunk.get()); // copy
}