vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

using abort_callback from ggml to stop llama computation

Browse source
This commit is contained in:
Michael Podvitskiy 2024-02-08 11:37:30 +01:00
parent 67fd33132f
commit e3ac833d3f
3 changed files with 22 additions and 3 deletions
Download patch file Download diff file Expand all files Collapse all files

1
ggml-backend.h
View file

@ -80,7 +80,6 @@ extern "C" {
// //
// CPU backend // CPU backend
// //
GGML_API ggml_backend_t ggml_backend_cpu_init(void); GGML_API ggml_backend_t ggml_backend_cpu_init(void);
GGML_API GGML_CALL bool ggml_backend_is_cpu (ggml_backend_t backend); GGML_API GGML_CALL bool ggml_backend_is_cpu (ggml_backend_t backend);

18
llama.cpp
View file

@ -1948,6 +1948,9 @@ struct llama_context {
std::vector<uint8_t> buf_compute_meta; std::vector<uint8_t> buf_compute_meta;
ggml_backend_sched_t sched = nullptr; ggml_backend_sched_t sched = nullptr;
ggml_abort_callback abort_callback = nullptr;
void * abort_callback_data = nullptr;
// input tensors // input tensors
ggml_backend_buffer_t buf_input = nullptr; ggml_backend_buffer_t buf_input = nullptr;
ggml_context * ctx_input = nullptr; ggml_context * ctx_input = nullptr;
@ -7847,6 +7850,7 @@ static void llama_graph_compute(
if (lctx.backend_cpu != nullptr) { if (lctx.backend_cpu != nullptr) {
ggml_backend_cpu_set_n_threads(lctx.backend_cpu, n_threads); ggml_backend_cpu_set_n_threads(lctx.backend_cpu, n_threads);
ggml_backend_cpu_set_abort_callback(lctx.backend_cpu, lctx.abort_callback, lctx.abort_callback_data);
} }
ggml_backend_sched_graph_compute(lctx.sched, gf); ggml_backend_sched_graph_compute(lctx.sched, gf);
@ -11644,6 +11648,8 @@ struct llama_context_params llama_context_default_params() {
/*.embedding =*/ false, /*.embedding =*/ false,
/*.offload_kqv =*/ true, /*.offload_kqv =*/ true,
/*.do_pooling =*/ true, /*.do_pooling =*/ true,
/*.abort_callback =*/ nullptr,
/*.abort_callback_data =*/ nullptr,
}; };
return result; return result;
@ -11835,8 +11841,11 @@ struct llama_context * llama_new_context_with_model(
LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base); LLAMA_LOG_INFO("%s: freq_base = %.1f\n", __func__, cparams.rope_freq_base);
LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale); LLAMA_LOG_INFO("%s: freq_scale = %g\n", __func__, cparams.rope_freq_scale);
ctx->rng = std::mt19937(params.seed); ctx->abort_callback = params.abort_callback;
ctx->logits_all = params.logits_all; ctx->abort_callback_data = params.abort_callback_data;
ctx->rng = std::mt19937(params.seed);
ctx->logits_all = params.logits_all;
const ggml_type type_k = params.type_k; const ggml_type type_k = params.type_k;
const ggml_type type_v = params.type_v; const ggml_type type_v = params.type_v;
@ -12809,6 +12818,11 @@ void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_
ctx->cparams.n_threads_batch = n_threads_batch; ctx->cparams.n_threads_batch = n_threads_batch;
} }
void llama_set_abort_callback(struct llama_context * ctx, bool (*abort_callback)(void * data), void * abort_callback_data) {
ctx->abort_callback = abort_callback;
ctx->abort_callback_data = abort_callback_data;
}
struct llama_batch llama_batch_get_one( struct llama_batch llama_batch_get_one(
llama_token * tokens, llama_token * tokens,
int32_t n_tokens, int32_t n_tokens,

6
llama.h
View file

@ -256,6 +256,9 @@ extern "C" {
bool embedding; // embedding mode only bool embedding; // embedding mode only
bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU bool offload_kqv; // whether to offload the KQV ops (including the KV cache) to GPU
bool do_pooling; // whether to pool (sum) embedding results by sequence id (ignored if no pooling layer) bool do_pooling; // whether to pool (sum) embedding results by sequence id (ignored if no pooling layer)
ggml_abort_callback abort_callback;
void * abort_callback_data;
}; };
// model quantization parameters // model quantization parameters
@ -661,6 +664,9 @@ extern "C" {
// n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens) // n_threads_batch is the number of threads used for prompt and batch processing (multiple tokens)
LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch); LLAMA_API void llama_set_n_threads(struct llama_context * ctx, uint32_t n_threads, uint32_t n_threads_batch);
// Set abort callback
LLAMA_API void llama_set_abort_callback(struct llama_context * ctx, ggml_abort_callback abort_callback, void * abort_callback_data);
// Token logits obtained from the last call to llama_eval() // Token logits obtained from the last call to llama_eval()
// The logits for the last token are stored in the last row // The logits for the last token are stored in the last row
// Logits for which llama_batch.logits[i] == 0 are undefined // Logits for which llama_batch.logits[i] == 0 are undefined