tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-28 20:25:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

graph : simplify attn input build for unified KV cache (#12381)

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-03-14 10:47:44 +02:00
committed by GitHub
parent 081bee8c64
commit c522ce4143
3 changed files with 53 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
src/llama-graph.cpp
View File

@ -1311,29 +1311,23 @@ ggml_tensor * llm_graph_context::build_attn(
return cur;
}
llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified(
bool causal,
bool swa) const {
llm_graph_input_attn_kv_unified * llm_graph_context::build_attn_inp_kv_unified() const {
const llama_kv_cache_unified * kv_self = static_cast<const llama_kv_cache_unified *>(memory);
auto inp = std::make_unique<llm_graph_input_attn_kv_unified>(hparams, cparams, kv_self);
const auto n_kv = kv_self->n;
inp->self_kq_mask = causal
? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD))
: ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
inp->self_kq_mask = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
//cb(inp->self_kq_mask, "KQ_mask", -1);
ggml_set_input(inp->self_kq_mask);
inp->self_kq_mask_cnv = cparams.flash_attn ? ggml_cast(ctx0, inp->self_kq_mask, GGML_TYPE_F16) : inp->self_kq_mask;
if (swa) {
if (hparams.n_swa_pattern > 1) {
GGML_ASSERT(hparams.n_swa > 0);
inp->self_kq_mask_swa = causal
? ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD))
: ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_tokens, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
inp->self_kq_mask_swa = ggml_new_tensor_2d(ctx0, GGML_TYPE_F32, n_kv, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
//cb(inp->self_kq_mask_swa, "KQ_mask_swa", -1);
ggml_set_input(inp->self_kq_mask_swa);