ikawrakow/ik_llama.cpp
1
0
Fork 0
mirror of https://github.com/ikawrakow/ik_llama.cpp.git synced 2026-04-30 19:31:48 +00:00
Code Issues Packages Projects Releases Wiki Activity

Minor

Browse Source
This commit is contained in:
Iwan Kawrakow
2024-08-28 18:31:27 +03:00
parent 4d10f4e0ba
commit 3b4fe65e1c
2 changed files with 4 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

1
ggml/src/ggml.c
View File

@@ -2181,7 +2181,6 @@ static inline float ggml_vec_add_f32_f32(const int n, const float * x, float * y
return max; return max;
} }
#else #else
// TODO add AVX2
static inline float ggml_vec_add_f32_f16(const int n, const ggml_half * x, float * y, float slope) { static inline float ggml_vec_add_f32_f16(const int n, const ggml_half * x, float * y, float slope) {
float max = -INFINITY; float max = -INFINITY;
for (int i = 0; i < n; ++i) { for (int i = 0; i < n; ++i) {

5
src/llama.cpp
View File

@@ -8692,8 +8692,9 @@ struct llm_build_context {
struct ggml_tensor * build_inp_KQ_mask(bool causal = true) { struct ggml_tensor * build_inp_KQ_mask(bool causal = true) {
auto nx = causal ? n_kv : n_tokens; 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 // 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 = !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; 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)); 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); cb(lctx.inp_KQ_mask, "KQ_mask", -1);
@@ -8705,6 +8706,8 @@ struct llm_build_context {
GGML_ASSERT(hparams.n_swa > 0); GGML_ASSERT(hparams.n_swa > 0);
auto nx = causal ? n_kv : n_tokens; 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 // 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 = !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; 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)); lctx.inp_KQ_mask_swa = ggml_new_tensor_2d(ctx0, type, nx, GGML_PAD(n_tokens, GGML_KQ_MASK_PAD));