public_git/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggml-org/whisper.cpp.git synced 2026-10-18 13:17:59 +02:00
Code Issues Packages Projects Releases Wiki Activity

talk-llama : sync llama.cpp

Browse Source
This commit is contained in:
Georgi Gerganov
2025-12-12 18:15:27 +02:00
parent 48cdc06e91
commit 179d8b1c9c
28 changed files with 2163 additions and 177 deletions
Download Patch File Download Diff File Expand all files Collapse all files
examples/talk-llama/llama-context.cpp
+11 -15
View File
@@ -1,5 +1,6 @@
#include "llama-context.h"
#include "llama-arch.h"
#include "llama-impl.h"
#include "llama-batch.h"
#include "llama-io.h"
@@ -92,14 +93,6 @@ llama_context::llama_context(
// with causal attention, the batch size is limited by the context size
cparams.n_batch = cparams.causal_attn ? std::min(cparams.n_ctx, params.n_batch) : params.n_batch;
// the batch has to be at least GGML_KQ_MASK_PAD because we will be padding the KQ_mask
// this is required by GPU kernels in order to avoid out-of-bounds accesses (e.g. ggml_flash_attn_ext)
// ref: https://github.com/ggerganov/llama.cpp/pull/5021
// TODO: this padding is not needed for the cache-less context so we should probably move it to llama_memory
if (cparams.n_batch < GGML_KQ_MASK_PAD) {
LLAMA_LOG_WARN("%s: n_batch is less than GGML_KQ_MASK_PAD - increasing to %d\n", __func__, GGML_KQ_MASK_PAD);
cparams.n_batch = GGML_KQ_MASK_PAD;
}
cparams.n_ubatch = std::min(cparams.n_batch, params.n_ubatch == 0 ? params.n_batch : params.n_ubatch);
cparams.op_offload = params.op_offload;
@@ -247,7 +240,10 @@ llama_context::llama_context(
LLAMA_LOG_DEBUG("%s: backend_ptrs.size() = %zu\n", __func__, backend_ptrs.size());
const size_t max_nodes = this->graph_max_nodes();
const uint32_t n_seqs = cparams.n_seq_max;
const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
const size_t max_nodes = this->graph_max_nodes(n_tokens);
LLAMA_LOG_DEBUG("%s: max_nodes = %zu\n", __func__, max_nodes);
@@ -299,9 +295,6 @@ llama_context::llama_context(
cross.v_embd.clear();
const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
// avoid reserving graphs with zero outputs - assume one output per sequence
n_outputs = n_seqs;
@@ -542,7 +535,7 @@ bool llama_context::memory_update(bool optimize) {
throw std::runtime_error("failed to initialize memory context");
}
const uint32_t n_seqs = cparams.kv_unified ? 1 : cparams.n_seq_max;
const uint32_t n_seqs = cparams.n_seq_max;
const uint32_t n_tokens = std::min(cparams.n_ctx, cparams.n_ubatch);
auto * gf = graph_reserve(n_tokens, n_seqs, n_tokens, mctx.get());
@@ -1248,7 +1241,7 @@ int llama_context::decode(const llama_batch & batch_inp) {
// make the outputs have the same order they had in the user-provided batch
// note: this is mostly relevant for recurrent models atm
if (!sorted_output) {
if (!sorted_output && n_outputs > 1) {
GGML_ASSERT((size_t) n_outputs == out_ids.size());
// TODO: is there something more efficient which also minimizes swaps?
@@ -1385,7 +1378,10 @@ void llama_context::output_reorder() {
// graph
//
uint32_t llama_context::graph_max_nodes() const {
uint32_t llama_context::graph_max_nodes(uint32_t n_tokens) const {
if (model.arch == LLM_ARCH_QWEN3NEXT) {
return std::max<uint32_t>(n_tokens * 40, 32u * model.n_tensors());
}
return std::max<uint32_t>(1024u, 8u*model.n_tensors());
}