root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-11-11 13:30:35 +00:00
Code Issues Actions 11 Packages Projects Releases Wiki Activity

squash! ggml : move rope type enum to ggml.h

Browse Source 
This commit removes the enum ggml_rope_type from ggml.h and replaces it
with a define (GGML_ROPE_TYPE_NEOX). This define is used in the code to
check if the mode is set to GPT-NeoX. Also the enum llama_rope_type has
been updated to reflect this change.
This commit is contained in:
Daniel Bevenius 2024-08-11 06:58:14 +02:00
parent 5e14dbf2ea
commit c9206f63be
Failed to extract signature
9 changed files with 11 additions and 15 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
ggml/include/ggml.h
View File

@ -244,6 +244,8 @@
#define GGML_EXIT_SUCCESS 0
#define GGML_EXIT_ABORTED 1
#define GGML_ROPE_TYPE_NEOX 2
#define GGUF_MAGIC "GGUF"
#define GGUF_VERSION 3
@ -437,12 +439,6 @@ extern "C" {
GGML_FTYPE_MOSTLY_Q4_0_8_8 = 27, // except 1d tensors
};
// Rotary Positional Embedding (RoPE) types
enum ggml_rope_type {
GGML_ROPE_TYPE_NORM = 0,
GGML_ROPE_TYPE_NEOX = 2,
};
// available tensor operations:
enum ggml_op {
GGML_OP_NONE = 0,
@ -1460,7 +1456,7 @@ extern "C" {
// rotary position embedding
// if mode & 1 == 1, skip n_past elements (NOT SUPPORTED)
// if mode & 2 == 1, GPT-NeoX style
// if mode & GGML_ROPE_TYPE_NEOX == 1, GPT-NeoX style
//
// b is an int32 vector with size a->ne[2], it contains the positions
GGML_API struct ggml_tensor * ggml_rope(

2
ggml/src/ggml-cann/aclnn_ops.cpp
View File

@ -2881,7 +2881,7 @@ void ggml_cann_rope(ggml_backend_cann_context& ctx, ggml_tensor* dst) {
ggml_rope_yarn_corr_dims(n_dims, n_ctx_orig, freq_base, beta_fast,
beta_slow, corr_dims);
const bool is_neox = mode & 2;
const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
// init cos/sin cache
ggml_cann_pool_alloc sin_allocator(

2
ggml/src/ggml-cuda/rope.cu
View File

@ -226,7 +226,7 @@ void ggml_cuda_op_rope(ggml_backend_cuda_context & ctx, ggml_tensor * dst) {
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
const bool is_neox = mode & 2;
const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
const int32_t * pos = (const int32_t *) src1_d;

2
ggml/src/ggml-metal.m
View File

@ -2313,7 +2313,7 @@ static enum ggml_status ggml_metal_graph_compute(
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
const bool is_neox = mode & 2;
const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
id<MTLComputePipelineState> pipeline = nil;

2
ggml/src/ggml-sycl/rope.cpp
View File

@ -226,7 +226,7 @@ void ggml_sycl_op_rope(
memcpy(&beta_fast, (int32_t *) dst->op_params + 9, sizeof(float));
memcpy(&beta_slow, (int32_t *) dst->op_params + 10, sizeof(float));
const bool is_neox = mode & 2;
const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
const int32_t * pos = (const int32_t *) src1_dd;

2
ggml/src/ggml-vulkan.cpp
View File

@ -4050,7 +4050,7 @@ static vk_pipeline ggml_vk_op_get_pipeline(ggml_backend_vk_context * ctx, const
case GGML_OP_ROPE:
{
const int mode = ((const int32_t *) dst->op_params)[2];
const bool is_neox = mode & 2;
const bool is_neox = mode & GGML_ROPE_TYPE_NEOX;
if (is_neox) {
if (src0->type == GGML_TYPE_F32 && dst->type == GGML_TYPE_F32) {

2
ggml/src/kompute-shaders/op_rope_f16.comp
View File

@ -11,7 +11,7 @@ void main() {
const uint i2 = gl_WorkGroupID.y;
const uint i1 = gl_WorkGroupID.x;
const bool is_neox = (pcs.mode & 2) != 0;
const bool is_neox = (pcs.mode & GGML_ROPE_TYPE_NEOX) != 0;
float corr_dims[2];
rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);

2
ggml/src/kompute-shaders/op_rope_f32.comp
View File

@ -11,7 +11,7 @@ void main() {
const uint i2 = gl_WorkGroupID.y;
const uint i1 = gl_WorkGroupID.x;
const bool is_neox = (pcs.mode & 2) != 0;
const bool is_neox = (pcs.mode & GGML_ROPE_TYPE_NEOX) != 0;
float corr_dims[2];
rope_yarn_corr_dims(pcs.n_dims, pcs.n_ctx_orig, pcs.freq_base, pcs.beta_fast, pcs.beta_slow, corr_dims);

2
include/llama.h
View File

@ -97,7 +97,7 @@ extern "C" {
enum llama_rope_type {
LLAMA_ROPE_TYPE_NONE = -1,
LLAMA_ROPE_TYPE_NORM = GGML_ROPE_TYPE_NORM,
LLAMA_ROPE_TYPE_NORM = 0,
LLAMA_ROPE_TYPE_NEOX = GGML_ROPE_TYPE_NEOX,
};