CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-14 11:07:10 +00:00
Code Issues Packages Projects Releases Wiki Activity

Merge branch 'master' into xsn/ggml_scale_bias

Browse Source
This commit is contained in:
Xuan Son Nguyen
2025-07-08 23:02:15 +02:00
parent 50f88fc4ca 6efcd65945
commit 7af3fd98a1
223 changed files with 9797 additions and 7507 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/include/ggml-backend.h
View File

@@ -339,7 +339,7 @@ extern "C" {
typedef bool (*ggml_backend_eval_callback)(int node_index, struct ggml_tensor * t1, struct ggml_tensor * t2, void * user_data);
// Compare the output of two backends
GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data);
GGML_API bool ggml_backend_compare_graph_backend(ggml_backend_t backend1, ggml_backend_t backend2, struct ggml_cgraph * graph, ggml_backend_eval_callback callback, void * user_data, struct ggml_tensor * test_node);
// Tensor initialization
GGML_API enum ggml_status ggml_backend_tensor_alloc(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, void * addr);

1
ggml/include/ggml-cpu.h
View File

@@ -134,6 +134,7 @@ extern "C" {
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
GGML_BACKEND_API void ggml_cpu_fp32_to_fp32(const float *, float *, int64_t);
GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);

50
ggml/include/ggml-kompute.h
View File

@@ -1,50 +0,0 @@
#pragma once
#include "ggml.h"
#include "ggml-backend.h"
#include <stdbool.h>
#include <stddef.h>
#include <stdint.h>
#ifdef __cplusplus
extern "C" {
#endif
#define GGML_KOMPUTE_MAX_DEVICES 16
struct ggml_vk_device {
int index;
int type; // same as VkPhysicalDeviceType
size_t heapSize;
const char * name;
const char * vendor;
int subgroupSize;
uint64_t bufferAlignment;
uint64_t maxAlloc;
};
struct ggml_vk_device * ggml_vk_available_devices(size_t memoryRequired, size_t * count);
bool ggml_vk_get_device(struct ggml_vk_device * device, size_t memoryRequired, const char * name);
bool ggml_vk_has_vulkan(void);
bool ggml_vk_has_device(void);
struct ggml_vk_device ggml_vk_current_device(void);
//
// backend API
//
// forward declaration
typedef struct ggml_backend * ggml_backend_t;
GGML_BACKEND_API ggml_backend_t ggml_backend_kompute_init(int device);
GGML_BACKEND_API bool ggml_backend_is_kompute(ggml_backend_t backend);
GGML_BACKEND_API ggml_backend_buffer_type_t ggml_backend_kompute_buffer_type(int device);
GGML_BACKEND_API ggml_backend_reg_t ggml_backend_kompute_reg(void);
#ifdef __cplusplus
}
#endif

191
ggml/include/ggml.h
View File

@@ -314,6 +314,13 @@
extern "C" {
#endif
// Function type used in fatal error callbacks
typedef void (*ggml_abort_callback_t)(const char * error_message);
// Set the abort callback (passing null will restore original abort functionality: printing a message to stdout)
// Returns the old callback for chaining
GGML_API ggml_abort_callback_t ggml_set_abort_callback(ggml_abort_callback_t callback);
GGML_NORETURN GGML_ATTRIBUTE_FORMAT(3, 4)
GGML_API void ggml_abort(const char * file, int line, const char * fmt, ...);
@@ -470,6 +477,7 @@ extern "C" {
GGML_OP_TRANSPOSE,
GGML_OP_GET_ROWS,
GGML_OP_GET_ROWS_BACK,
GGML_OP_SET_ROWS,
GGML_OP_DIAG,
GGML_OP_DIAG_MASK_INF,
GGML_OP_DIAG_MASK_ZERO,
@@ -481,12 +489,13 @@ extern "C" {
GGML_OP_CONV_TRANSPOSE_1D,
GGML_OP_IM2COL,
GGML_OP_IM2COL_BACK,
GGML_OP_CONV_2D,
GGML_OP_CONV_2D_DW,
GGML_OP_CONV_TRANSPOSE_2D,
GGML_OP_POOL_1D,
GGML_OP_POOL_2D,
GGML_OP_POOL_2D_BACK,
GGML_OP_UPSCALE, // nearest interpolate
GGML_OP_UPSCALE,
GGML_OP_PAD,
GGML_OP_PAD_REFLECT_1D,
GGML_OP_ROLL,
@@ -519,6 +528,8 @@ extern "C" {
GGML_OP_CROSS_ENTROPY_LOSS_BACK,
GGML_OP_OPT_STEP_ADAMW,
GGML_OP_GLU,
GGML_OP_COUNT,
};
@@ -542,6 +553,16 @@ extern "C" {
GGML_UNARY_OP_COUNT,
};
enum ggml_glu_op {
GGML_GLU_OP_REGLU,
GGML_GLU_OP_GEGLU,
GGML_GLU_OP_SWIGLU,
GGML_GLU_OP_GEGLU_ERF,
GGML_GLU_OP_GEGLU_QUICK,
GGML_GLU_OP_COUNT,
};
enum ggml_object_type {
GGML_OBJECT_TYPE_TENSOR,
GGML_OBJECT_TYPE_GRAPH,
@@ -627,6 +648,9 @@ extern "C" {
// misc
GGML_API const char * ggml_version(void);
GGML_API const char * ggml_commit(void);
GGML_API void ggml_time_init(void); // call this once at the beginning of the program
GGML_API int64_t ggml_time_ms(void);
GGML_API int64_t ggml_time_us(void);
@@ -657,6 +681,7 @@ extern "C" {
GGML_API const char * ggml_op_symbol(enum ggml_op op);
GGML_API const char * ggml_unary_op_name(enum ggml_unary_op op);
GGML_API const char * ggml_glu_op_name(enum ggml_glu_op op);
GGML_API const char * ggml_op_desc(const struct ggml_tensor * t); // unary or op name
GGML_API size_t ggml_element_size(const struct ggml_tensor * tensor);
@@ -687,6 +712,9 @@ extern "C" {
// true for tensor that is stored in memory as CxWxHxN and has been permuted to WxHxCxN
GGML_API bool ggml_is_contiguous_channels(const struct ggml_tensor * tensor);
// true if the elements in dimension 0 are contiguous, or there is just 1 block of elements
GGML_API bool ggml_is_contiguous_rows(const struct ggml_tensor * tensor);
GGML_API bool ggml_are_same_shape (const struct ggml_tensor * t0, const struct ggml_tensor * t1);
GGML_API bool ggml_are_same_stride(const struct ggml_tensor * t0, const struct ggml_tensor * t1);
@@ -758,6 +786,7 @@ extern "C" {
GGML_API void ggml_unravel_index(const struct ggml_tensor * tensor, int64_t i, int64_t * i0, int64_t * i1, int64_t * i2, int64_t * i3);
GGML_API enum ggml_unary_op ggml_get_unary_op(const struct ggml_tensor * tensor);
GGML_API enum ggml_glu_op ggml_get_glu_op(const struct ggml_tensor * tensor);
GGML_API void * ggml_get_data (const struct ggml_tensor * tensor);
GGML_API float * ggml_get_data_f32(const struct ggml_tensor * tensor);
@@ -1086,6 +1115,89 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);
// gated linear unit ops
// A: n columns, r rows,
// result is n / 2 columns, r rows,
// expects gate in second half of row, unless swapped is true
GGML_API struct ggml_tensor * ggml_glu(
struct ggml_context * ctx,
struct ggml_tensor * a,
enum ggml_glu_op op,
bool swapped);
GGML_API struct ggml_tensor * ggml_reglu(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_reglu_swapped(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_geglu(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_geglu_swapped(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_swiglu(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_swiglu_swapped(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_geglu_erf(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_geglu_erf_swapped(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_geglu_quick(
struct ggml_context * ctx,
struct ggml_tensor * a);
GGML_API struct ggml_tensor * ggml_geglu_quick_swapped(
struct ggml_context * ctx,
struct ggml_tensor * a);
// A: n columns, r rows,
// B: n columns, r rows,
GGML_API struct ggml_tensor * ggml_glu_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b,
enum ggml_glu_op op);
GGML_API struct ggml_tensor * ggml_reglu_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_geglu_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_swiglu_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_geglu_erf_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
GGML_API struct ggml_tensor * ggml_geglu_quick_split(
struct ggml_context * ctx,
struct ggml_tensor * a,
struct ggml_tensor * b);
// normalize along rows
GGML_API struct ggml_tensor * ggml_norm(
struct ggml_context * ctx,
@@ -1388,6 +1500,23 @@ extern "C" {
struct ggml_tensor * b, // row indices
struct ggml_tensor * c); // data for ggml_get_rows, only used for its shape
// a TD [n_embd, ne1, ne2, ne3]
// b TS [n_embd, n_rows, ne02, ne03] | ne02 == ne2, ne03 == ne3
// c I64 [n_rows, ne11, ne12, 1] | c[i] in [0, ne1)
//
// undefined behavior if destination rows overlap
//
// broadcast:
// ne2 % ne11 == 0
// ne3 % ne12 == 0
//
// return view(a)
GGML_API struct ggml_tensor * ggml_set_rows(
struct ggml_context * ctx,
struct ggml_tensor * a, // destination
struct ggml_tensor * b, // source
struct ggml_tensor * c); // row indices
GGML_API struct ggml_tensor * ggml_diag(
struct ggml_context * ctx,
struct ggml_tensor * a);
@@ -1425,8 +1554,14 @@ extern "C" {
struct ggml_context * ctx,
struct ggml_tensor * a);
// a [ne0, ne01, ne02, ne03]
// mask [ne0, ne11, ne12, ne13] | ne11 >= ne01, F16 or F32, optional
//
// broadcast:
// ne02 % ne12 == 0
// ne03 % ne13 == 0
//
// fused soft_max(a*scale + mask*(ALiBi slope))
// mask is optional
// max_bias = 0.0f for no ALiBi
GGML_API struct ggml_tensor * ggml_soft_max_ext(
struct ggml_context * ctx,
@@ -1736,6 +1871,17 @@ extern "C" {
struct ggml_tensor * b,
int stride);
GGML_API struct ggml_tensor * ggml_conv_2d_direct(
struct ggml_context * ctx,
struct ggml_tensor * a, // convolution kernel [KW, KH, IC, OC]
struct ggml_tensor * b, // input data [W, H, C, N]
int s0, // stride dimension 0
int s1, // stride dimension 1
int p0, // padding dimension 0
int p1, // padding dimension 1
int d0, // dilation dimension 0
int d1); // dilation dimension 1
enum ggml_op_pool {
GGML_OP_POOL_MAX,
GGML_OP_POOL_AVG,
@@ -1778,6 +1924,12 @@ extern "C" {
enum ggml_scale_mode {
GGML_SCALE_MODE_NEAREST = 0,
GGML_SCALE_MODE_BILINEAR = 1,
GGML_SCALE_MODE_COUNT
};
enum ggml_scale_flag {
GGML_SCALE_FLAG_ALIGN_CORNERS = (1 << 8)
};
// interpolate
@@ -1790,14 +1942,26 @@ extern "C" {
// interpolate
// interpolate scale to specified dimensions
GGML_API struct ggml_tensor * ggml_upscale_ext(
GGML_DEPRECATED(GGML_API struct ggml_tensor * ggml_upscale_ext(
struct ggml_context * ctx,
struct ggml_tensor * a,
int ne0,
int ne1,
int ne2,
int ne3,
enum ggml_scale_mode mode);
enum ggml_scale_mode mode),
"use ggml_interpolate instead");
// Up- or downsamples the input to the specified size.
// 2D scale modes (eg. bilinear) are applied to the first two dimensions.
GGML_API struct ggml_tensor * ggml_interpolate(
struct ggml_context * ctx,
struct ggml_tensor * a,
int64_t ne0,
int64_t ne1,
int64_t ne2,
int64_t ne3,
uint32_t mode); // ggml_scale_mode [ | ggml_scale_flag...]
// pad each dimension with zeros: [x, ..., x] -> [x, ..., x, 0, ..., 0]
GGML_API struct ggml_tensor * ggml_pad(
@@ -1860,11 +2024,17 @@ extern "C" {
#define GGML_KQ_MASK_PAD 64
// q: [n_embd_k, n_batch, n_head, 1]
// k: [n_embd_k, n_kv, n_head_kv, 1]
// v: [n_embd_v, n_kv, n_head_kv, 1] !! not transposed !!
// mask: [n_kv, n_batch_pad, 1, 1] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
// res: [n_embd_v, n_head, n_batch, 1] !! permuted !!
// q: [n_embd_k, n_batch, n_head, ne3 ]
// k: [n_embd_k, n_kv, n_head_kv, ne3 ]
// v: [n_embd_v, n_kv, n_head_kv, ne3 ] !! not transposed !!
// mask: [n_kv, n_batch_pad, ne32, ne33] !! n_batch_pad = GGML_PAD(n_batch, GGML_KQ_MASK_PAD) !!
// res: [n_embd_v, n_head, n_batch, ne3 ] !! permuted !!
//
// broadcast:
// n_head % n_head_kv == 0
// n_head % ne32 == 0
// ne3 % ne33 == 0
//
GGML_API struct ggml_tensor * ggml_flash_attn_ext(
struct ggml_context * ctx,
struct ggml_tensor * q,
@@ -1903,7 +2073,8 @@ extern "C" {
struct ggml_tensor * dt,
struct ggml_tensor * A,
struct ggml_tensor * B,
struct ggml_tensor * C);
struct ggml_tensor * C,
struct ggml_tensor * ids);
// partition into non-overlapping windows with padding if needed
// example: