Superminaren/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2026-09-27 21:46:57 +02:00
Code Issues Packages Projects Releases Wiki Activity

llama: add llama_batch_ext (#24669)

Browse Source 
* (wip) add llama_batch_ext

* wip

* updated design

* updated impl

* change signature

* unused var

* demo common_prompt_batch_decode

* fix pos

* tmp disable test-batch-alloc

* fix compat

* nits: add const

* no more pos_max

* add comment about llama_batch_ext_set_embd_state

* handle n_embd_out properly

* rename api --> embd_token

* llama_embd

* stub llama_batch_ext_set_embd_state

* support both token + embd + state in batch

* llama_batch_ext_add_embd

* upstream some changes

* nits

* fix test-batch-alloc

* add test for compat
This commit is contained in:
Xuan-Son Nguyen
2026-09-24 16:25:07 +02:00
committed by GitHub
parent 308883b335
commit fc343a84bb
10 changed files with 1210 additions and 262 deletions
Download Patch File Download Diff File Expand all files Collapse all files
common/common.cpp
+30 -7
View File
@@ -2198,9 +2198,28 @@ bool common_replay_last_token(struct llama_context * ctx, llama_token last_token
return true;
}
llama_batch_ext_ptr common_batch_ext_get_one(llama_context * ctx, const llama_tokens & tokens) {
llama_batch_ext_ptr batch(llama_batch_ext_init(ctx));
auto mem = llama_get_memory(ctx);
llama_pos pos = mem ? llama_memory_seq_pos_max(mem, 0) + 1 : 0;
for (size_t i = 0; i < tokens.size(); ++i) {
const int32_t idx = llama_batch_ext_add_token(batch.get(), 0, tokens[i]);
llama_batch_ext_set_pos(batch.get(), idx, &pos);
pos++;
}
if (!tokens.empty()) {
llama_batch_ext_set_output_logits(batch.get(), (int32_t) tokens.size() - 1, true);
}
return batch;
}
bool common_prompt_batch_decode(
struct llama_context * ctx,
const std::vector<llama_token> & all_tokens,
const llama_tokens & all_tokens,
int n_new,
int & n_past,
int n_batch,
@@ -2221,7 +2240,9 @@ bool common_prompt_batch_decode(
// Memory implementations in recurrent/hybrid models don't support removing tokens from their
// memory, so we can't just remove the last token from the memory and replay the last token which
// is the reason for this logic.
if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(all_tokens.data() + offset), n_tokens_before_last))) {
llama_tokens prefix_tokens(all_tokens.begin() + offset, all_tokens.begin() + offset + n_tokens_before_last);
llama_batch_ext_ptr batch_prefix = common_batch_ext_get_one(ctx, prefix_tokens);
if (llama_process(ctx, LLAMA_PROCESS_TYPE_DECODE, batch_prefix.get())) {
COM_ERR("%s", "failed to eval\n");
return false;
}
@@ -2231,17 +2252,19 @@ bool common_prompt_batch_decode(
COM_INF("saved session before last token to %s, n_new = %zu\n", state_path.data(), all_tokens.size());
llama_token last_token = all_tokens.back();
llama_batch batch = llama_batch_get_one(&last_token, 1);
int32_t pos = n_past;
batch.pos = &pos;
llama_batch_ext_ptr batch_last = common_batch_ext_get_one(ctx, { last_token });
llama_pos pos = n_past;
llama_batch_ext_set_pos(batch_last.get(), 0, &pos);
if (llama_decode(ctx, batch)) {
if (llama_process(ctx, LLAMA_PROCESS_TYPE_DECODE, batch_last.get())) {
COM_ERR("%s", "failed to eval last token\n");
return false;
}
n_past++;
} else {
if (llama_decode(ctx, llama_batch_get_one(const_cast<llama_token*>(all_tokens.data() + offset), n_new))) {
llama_tokens new_tokens(all_tokens.begin() + offset, all_tokens.begin() + offset + n_new);
llama_batch_ext_ptr batch = common_batch_ext_get_one(ctx, new_tokens);
if (llama_process(ctx, LLAMA_PROCESS_TYPE_DECODE, batch.get())) {
COM_ERR("%s", "failed to eval\n");
return false;
}