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

llama : use zero value of n_swa to distinguish Phi-4 from other PHI3 models

Browse source
This commit is contained in:
Stanisław Szymczyk 2024-12-14 12:00:19 +01:00
parent c7fdbd3735
commit 046c0d77a9
1 changed files with 2 additions and 2 deletions
Download patch file Download diff file Expand all files Collapse all files

4
src/llama.cpp
View file

@ -5807,7 +5807,7 @@ static void llm_load_hparams(
hparams.n_swa = 131072; hparams.n_swa = 131072;
} }
bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false); bool found_swa = ml.get_key(LLM_KV_ATTENTION_SLIDING_WINDOW, hparams.n_swa, false);
if (!found_swa && hparams.n_swa == 0 && model.name != "Phi 4") { if (!found_swa && hparams.n_swa == 0) {
throw std::runtime_error("invalid value for sliding_window"); throw std::runtime_error("invalid value for sliding_window");
} }
} break; } break;
@ -12840,7 +12840,7 @@ struct llm_build_context {
// KQ_mask (mask for 1 head, it will be broadcasted to all heads) // KQ_mask (mask for 1 head, it will be broadcasted to all heads)
struct ggml_tensor * KQ_mask = nullptr; struct ggml_tensor * KQ_mask = nullptr;
if (model.name == "Phi 4") { if (hparams.n_swa == 0) {
// Phi-4 doesn't use sliding window attention // Phi-4 doesn't use sliding window attention
KQ_mask = build_inp_KQ_mask(); KQ_mask = build_inp_KQ_mask();
} else { } else {