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

Merge branch 'master' into compilade/bitnet-ternary

Browse source
This commit is contained in:
Francis Couture-Harpin 2024-08-13 17:36:09 -04:00
commit 82b240406d
22 changed files with 75 additions and 58 deletions
Download patch file Download diff file Expand all files Collapse all files

6
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
@ -1455,8 +1457,8 @@ extern "C" {
struct ggml_tensor * b);
// rotary position embedding
// if mode & 1 == 1, skip n_past elements (NOT SUPPORTED)
// if mode & 2 == 1, GPT-NeoX style
// if (mode & 1) - skip n_past elements (NOT SUPPORTED)
// if (mode & GGML_ROPE_TYPE_NEOX) - 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

@ -4053,7 +4053,7 @@ 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) {

6
ggml/src/ggml.c
View file

@ -14131,7 +14131,7 @@ static void ggml_compute_forward_rope_f32(
float corr_dims[2];
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;
const float * freq_factors = NULL;
if (src2 != NULL) {
@ -14256,7 +14256,7 @@ static void ggml_compute_forward_rope_f16(
float corr_dims[2];
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;
const float * freq_factors = NULL;
if (src2 != NULL) {
@ -21168,7 +21168,7 @@ struct gguf_context * gguf_init_from_file(const char * fname, struct gguf_init_p
(int64_t) info->ne[2] *
(int64_t) info->ne[3];
if (ne % ggml_blck_size(info->type) != 0) {
if (ggml_blck_size(info->type) == 0 || ne % ggml_blck_size(info->type) != 0) {
fprintf(stderr, "%s: tensor '%s' of type %d (%s) number of elements (%" PRId64 ") is not a multiple of block size (%" PRId64 ")\n",
__func__, info->name.data, (int) info->type, ggml_type_name(info->type), ne, ggml_blck_size(info->type));
fclose(file);

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
ggml/src/kompute-shaders/rope_common.comp
View file

@ -1,5 +1,7 @@
#include "common.comp"
#define GGML_ROPE_TYPE_NEOX 2
// TODO: use a local size of 32 or more (Metal uses 1024)
layout(local_size_x = 1) in;