From 3b4fe65e1cc1e7f8172cb957dcb01b207df373d8 Mon Sep 17 00:00:00 2001
From: Iwan Kawrakow <iwan.kawrakow@gmail.com>
Date: Wed, 28 Aug 2024 18:31:27 +0300
Subject: [PATCH] Minor

---
 ggml/src/ggml.c | 1 -
 src/llama.cpp   | 5 ++++-
 2 files changed, 4 insertions(+), 2 deletions(-)

diff --git a/ggml/src/ggml.c b/ggml/src/ggml.c
index e1cf168b..939f5215 100644
--- a/ggml/src/ggml.c
+++ b/ggml/src/ggml.c
@@ -2181,7 +2181,6 @@ static inline float ggml_vec_add_f32_f32(const int n, const float * x, float * y
     return max;
 }
 #else
-// TODO add AVX2
 static inline float ggml_vec_add_f32_f16(const int n, const ggml_half * x, float * y, float slope) {
     float max = -INFINITY;
     for (int i = 0; i < n; ++i) {
diff --git a/src/llama.cpp b/src/llama.cpp
index 59370600..ad3febef 100644
--- a/src/llama.cpp
+++ b/src/llama.cpp
@@ -8692,8 +8692,9 @@ struct llm_build_context {
     struct ggml_tensor * build_inp_KQ_mask(bool causal = true) {
         auto nx = causal ? n_kv : n_tokens;
         // Note: we only use a binary mask when nx%32 == 0 because otherwise the CUDA implementation becomes way more messy
+        //bool can_be_binary = binary_kq && !lctx.is_encoding && !flash_attn && !hparams.use_alibi && nx%32 == 0;
+        //auto type = can_be_binary ? GGML_TYPE_I32 : flash_attn ? GGML_TYPE_F16 : GGML_TYPE_F32;
         auto type = !lctx.is_encoding ? !binary_kq || flash_attn || hparams.use_alibi || (nx%32 != 0) ? GGML_TYPE_F16 : GGML_TYPE_I32 : GGML_TYPE_F32;
-        //auto type = flash_attn || hparams.use_alibi || (nx%32 != 0) ? GGML_TYPE_F16 : GGML_TYPE_I32;
         if (type == GGML_TYPE_I32) nx /= 32;
         lctx.inp_KQ_mask = ggml_new_tensor_2d(ctx0, type, nx, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));
         cb(lctx.inp_KQ_mask, "KQ_mask", -1);
@@ -8705,6 +8706,8 @@ struct llm_build_context {
         GGML_ASSERT(hparams.n_swa > 0);
         auto nx = causal ? n_kv : n_tokens;
         // Note: we only use a binary mask when nx%32 == 0 because otherwise the CUDA implementation becomes way more messy
+        //bool can_be_binary = binary_kq && !lctx.is_encoding && !flash_attn && !hparams.use_alibi && nx%32 == 0;
+        //auto type = can_be_binary ? GGML_TYPE_I32 : flash_attn ? GGML_TYPE_F16 : GGML_TYPE_F32;
         auto type = !lctx.is_encoding ? !binary_kq || flash_attn || hparams.use_alibi || (nx%32 != 0) ? GGML_TYPE_F16 : GGML_TYPE_I32 : GGML_TYPE_F32;
         if (type == GGML_TYPE_I32) nx /= 32;
         lctx.inp_KQ_mask_swa = ggml_new_tensor_2d(ctx0, type, nx, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));