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

ggml : full ALiBi support

Browse source
This commit is contained in:
Georgi Gerganov 2024-05-10 10:40:33 +03:00
parent d11afd6652
commit 7fdca3348c
No known key found for this signature in database
GPG key ID: 449E073F9DC10735
10 changed files with 82 additions and 680 deletions
Download patch file Download diff file Expand all files Collapse all files

15
ggml.h
View file

@ -468,7 +468,6 @@ extern "C" {
GGML_OP_SOFT_MAX_BACK,
GGML_OP_ROPE,
GGML_OP_ROPE_BACK,
GGML_OP_ALIBI,
GGML_OP_CLAMP,
GGML_OP_CONV_TRANSPOSE_1D,
GGML_OP_IM2COL,
@ -1428,15 +1427,13 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);
// fused soft_max(a*scale + mask + pos[i]*(ALiBi slope))
// fused soft_max(a*scale + mask*(ALiBi slope))
// mask is optional
// pos is required when max_bias > 0.0f
// max_bias = 0.0f for no ALiBi
GGML_API struct ggml_tensor * ggml_soft_max_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * mask,
struct ggml_tensor * pos,
float scale,
float max_bias);
@ -1538,16 +1535,6 @@ extern "C" {
float xpos_base,
bool xpos_down);
// alibi position embedding
// in-place, returns view(a)
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_alibi(
struct ggml_context * ctx,
struct ggml_tensor * a,
int n_past,
int n_head,
float bias_max),
"use ggml_soft_max_ext instead (will be removed in Mar 2024)");
// clamp
// in-place, returns view(a)
GGML_API struct ggml_tensor * ggml_clamp(