root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-25 10:54:36 +00:00
Code Issues Actions 9 Packages Projects Releases Wiki Activity

ggml : reuse quantum structs across backends (#5943)

Browse Source 
* ggml : reuse quant blocks across backends

ggml-ci

* ggml : define helper constants only for CUDA and SYCL

ggml-ci

* ggml : define helper quantum constants for SYCL

ggml-ci
This commit is contained in:
Georgi Gerganov 2024-03-12 14:27:20 +02:00 committed by GitHub
parent 184215e783
commit 8030da7afe
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
7 changed files with 470 additions and 881 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

409
ggml-common.h
View File

@ -1,4 +1,408 @@
#pragma once
#ifndef GGML_COMMON_DECL
#if defined(GGML_COMMON_DECL_C)
#include <stdint.h>
typedef uint16_t ggml_half;
typedef uint32_t ggml_half2;
#define GGML_COMMON_AGGR
#define GGML_COMMON_DECL
#elif defined(GGML_COMMON_DECL_METAL)
#include <metal_stdlib>
typedef half ggml_half;
typedef half2 ggml_half2;
#define GGML_COMMON_AGGR
#define GGML_COMMON_DECL
#elif defined(GGML_COMMON_DECL_CUDA)
#include <cuda_fp16.h>
#include <cstdint>
typedef half ggml_half;
typedef half2 ggml_half2;
#define GGML_COMMON_AGGR data
#define GGML_COMMON_DECL
#elif defined(GGML_COMMON_DECL_HIP)
#include <hip/hip_fp16.h>
#include <cstdint>
typedef half ggml_half;
typedef half2 ggml_half2;
#define GGML_COMMON_AGGR data
#define GGML_COMMON_DECL
#elif defined(GGML_COMMON_DECL_SYCL)
#include <sycl/half_type.hpp>
#include <cstdint>
typedef sycl::half ggml_half;
typedef sycl::half2 ggml_half2;
#define GGML_COMMON_AGGR data
#define GGML_COMMON_DECL
#endif
#if defined(GGML_COMMON_DECL)
#ifndef __cplusplus
#ifndef static_assert
#if defined(__STDC_VERSION__) && (__STDC_VERSION__ >= 201100L)
#define static_assert(cond, msg) _Static_assert(cond, msg)
#else
#define static_assert(cond, msg) struct global_scope_noop_trick
#endif
#endif
#endif // __cplusplus
// QK = number of values after dequantization
// QK_K = super-block size
#ifdef GGML_QKK_64
#define QK_K 64
#define K_SCALE_SIZE 4
#else
#define QK_K 256
#define K_SCALE_SIZE 12
#endif // GGML_QKK_64
#if defined(GGML_COMMON_DECL_CUDA) || defined(GGML_COMMON_DECL_HIP) || defined(GGML_COMMON_DECL_SYCL)
// QR = QK / number of values before dequantization
// QI = number of 32 bit integers before dequantization
#define QI4_0 (QK4_0 / (4 * QR4_0))
#define QR4_0 2
#define QI4_1 (QK4_1 / (4 * QR4_1))
#define QR4_1 2
#define QI5_0 (QK5_0 / (4 * QR5_0))
#define QR5_0 2
#define QI5_1 (QK5_1 / (4 * QR5_1))
#define QR5_1 2
#define QI8_0 (QK8_0 / (4 * QR8_0))
#define QR8_0 1
#define QI8_1 (QK8_1 / (4 * QR8_1))
#define QR8_1 1
#define QI2_K (QK_K / (4*QR2_K))
#define QR2_K 4
#define QI3_K (QK_K / (4*QR3_K))
#define QR3_K 4
#define QI4_K (QK_K / (4*QR4_K))
#define QR4_K 2
#define QI5_K (QK_K / (4*QR5_K))
#define QR5_K 2
#define QI6_K (QK_K / (4*QR6_K))
#define QR6_K 2
#define QI2_XXS (QK_K / (4*QR2_XXS))
#define QR2_XXS 8
#define QI2_XS (QK_K / (4*QR2_XS))
#define QR2_XS 8
#define QI2_S (QK_K / (4*QR2_S))
#define QR2_S 8
#define QI3_XXS (QK_K / (4*QR3_XXS))
#define QR3_XXS 8
#define QI3_XS (QK_K / (4*QR3_XS))
#define QR3_XS 8
#define QI1_S (QK_K / (4*QR1_S))
#define QR1_S 8
#define QI4_NL (QK4_NL / (4*QR4_NL))
#define QR4_NL 2
#if QK_K == 64
#define QI4_XS QI4_NL
#define QR4_XS QR4_NL
#else
#define QI4_XS (QK_K / (4*QR4_XS))
#define QR4_XS 8
#endif
#endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
#define QK4_0 32
typedef struct {
ggml_half d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
static_assert(sizeof(block_q4_0) == sizeof(ggml_half) + QK4_0 / 2, "wrong q4_0 block size/padding");
#define QK4_1 32
typedef struct {
union {
struct {
ggml_half d; // delta
ggml_half m; // min
} GGML_COMMON_AGGR;
ggml_half2 dm;
};
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_half) + QK4_1 / 2, "wrong q4_1 block size/padding");
#define QK5_0 32
typedef struct {
ggml_half d; // delta
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_0 / 2]; // nibbles / quants
} block_q5_0;
static_assert(sizeof(block_q5_0) == sizeof(ggml_half) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
#define QK5_1 32
typedef struct {
union {
struct {
ggml_half d; // delta
ggml_half m; // min
} GGML_COMMON_AGGR;
ggml_half2 dm;
};
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_1 / 2]; // nibbles / quants
} block_q5_1;
static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_half) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
#define QK8_0 32
typedef struct {
ggml_half d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
static_assert(sizeof(block_q8_0) == sizeof(ggml_half) + QK8_0, "wrong q8_0 block size/padding");
#define QK8_1 32
typedef struct {
union {
struct {
ggml_half d; // delta
ggml_half s; // d * sum(qs[i])
} GGML_COMMON_AGGR;
ggml_half2 ds;
};
int8_t qs[QK8_1]; // quants
} block_q8_1;
static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_half) + QK8_1, "wrong q8_1 block size/padding");
//
// Super-block quantization structures
//
// 2-bit quantization
// weight is represented as x = a * q + b
// 16 blocks of 16 elements each
// Effectively 2.625 bits per weight
typedef struct {
uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
uint8_t qs[QK_K/4]; // quants
union {
struct {
ggml_half d; // super-block scale for quantized scales
ggml_half dmin; // super-block scale for quantized mins
} GGML_COMMON_AGGR;
ggml_half2 dm;
};
} block_q2_K;
static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_half) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
// 3-bit quantization
// weight is represented as x = a * q
// 16 blocks of 16 elements each
// Effectively 3.4375 bits per weight
#ifdef GGML_QKK_64
typedef struct {
uint8_t hmask[QK_K/8]; // quants - high bit
uint8_t qs[QK_K/4]; // quants - low 2 bits
uint8_t scales[2];
ggml_half d; // super-block scale
} block_q3_K;
static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
#else
typedef struct {
uint8_t hmask[QK_K/8]; // quants - high bit
uint8_t qs[QK_K/4]; // quants - low 2 bits
uint8_t scales[12]; // scales, quantized with 6 bits
ggml_half d; // super-block scale
} block_q3_K;
static_assert(sizeof(block_q3_K) == sizeof(ggml_half) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
#endif
// 4-bit quantization
// 8 blocks of 32 elements each
// weight is represented as x = a * q + b
// Effectively 4.5 bits per weight
#ifdef GGML_QKK_64
typedef struct {
ggml_half d[2]; // super-block scales/mins
uint8_t scales[2]; // 4-bit block scales/mins
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + QK_K/2 + 2, "wrong q4_K block size/padding");
#else
typedef struct {
union {
struct {
ggml_half d; // super-block scale for quantized scales
ggml_half dmin; // super-block scale for quantized mins
} GGML_COMMON_AGGR;
ggml_half2 dm;
};
uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
#endif
// 5-bit quantization
// 8 blocks of 32 elements each
// weight is represented as x = a * q + b
// Effectively 5.5 bits per weight
#ifdef GGML_QKK_64
typedef struct {
ggml_half d; // super-block scale
int8_t scales[QK_K/16]; // 8-bit block scales
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == sizeof(ggml_half) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
#else
typedef struct {
union {
struct {
ggml_half d; // super-block scale for quantized scales
ggml_half dmin; // super-block scale for quantized mins
} GGML_COMMON_AGGR;
ggml_half2 dm;
};
uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_half) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
#endif
// 6-bit quantization
// weight is represented as x = a * q
// 16 blocks of 16 elements each
// Effectively 6.5625 bits per weight
typedef struct {
uint8_t ql[QK_K/2]; // quants, lower 4 bits
uint8_t qh[QK_K/4]; // quants, upper 2 bits
int8_t scales[QK_K/16]; // scales, quantized with 8 bits
ggml_half d; // super-block scale
} block_q6_K;
static_assert(sizeof(block_q6_K) == sizeof(ggml_half) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
// This is only used for intermediate quantization and dot products
typedef struct {
float d; // delta
int8_t qs[QK_K]; // quants
int16_t bsums[QK_K/16]; // sum of quants in groups of 16
} block_q8_K;
static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
// (Almost) "true" 2-bit quantization.
// Due to the need to use blocks as per ggml design, it ends up using
// 2.0625 bpw because of the 16-bit scale for each block of 256.
typedef struct {
ggml_half d;
uint16_t qs[QK_K/8];
} block_iq2_xxs;
static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
// 2.3125 bpw quants
typedef struct {
ggml_half d;
uint16_t qs[QK_K/8];
uint8_t scales[QK_K/32];
} block_iq2_xs;
static_assert(sizeof(block_iq2_xs) == sizeof(ggml_half) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
// 2.5625 bpw quants
typedef struct {
ggml_half d;
uint8_t qs[QK_K/4];
uint8_t qh[QK_K/32];
uint8_t scales[QK_K/32];
} block_iq2_s;
static_assert(sizeof(block_iq2_s) == sizeof(ggml_half) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
// (Almost) "true" 3-bit quantization.
// Due to the need to use blocks as per ggml design, it ends up using
// 3.0625 bpw because of the 16-bit scale for each block of 256.
typedef struct {
ggml_half d;
uint8_t qs[3*QK_K/8];
} block_iq3_xxs;
static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_half) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
// 3.4375 bpw
#if QK_K == 64
#define IQ3S_N_SCALE 2
#else
#define IQ3S_N_SCALE QK_K/64
#endif
typedef struct {
ggml_half d;
uint8_t qs[QK_K/4];
uint8_t qh[QK_K/32];
uint8_t signs[QK_K/8];
uint8_t scales[IQ3S_N_SCALE];
} block_iq3_s;
static_assert(sizeof(block_iq3_s) == sizeof(ggml_half) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
typedef struct {
ggml_half d;
uint8_t qs[QK_K/8];
uint16_t qh[QK_K/32];
} block_iq1_s;
static_assert(sizeof(block_iq1_s) == sizeof(ggml_half) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
// Non-linear quants
#define QK4_NL 32
typedef struct {
ggml_half d;
uint8_t qs[QK4_NL/2];
} block_iq4_nl;
static_assert(sizeof(block_iq4_nl) == sizeof(ggml_half) + QK4_NL/2, "wrong iq4_nl block size/padding");
#if QK_K == 64
#define block_iq4_xs block_iq4_nl
#else
typedef struct {
ggml_half d;
uint16_t scales_h;
uint8_t scales_l[QK_K/64];
uint8_t qs[QK_K/2];
} block_iq4_xs;
static_assert(sizeof(block_iq4_xs) == sizeof(ggml_half) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
#endif
#endif // GGML_COMMON_DECL
#endif // GGML_COMMON_DECL
////////////////////////////////////////////////////////////////////////////////
#ifndef GGML_COMMON_IMPL
#if defined(GGML_COMMON_IMPL_C)
#include <stdint.h>
@ -14,7 +418,7 @@
#define GGML_TABLE_END() };
#define GGML_COMMON_IMPL
#elif defined(GGML_COMMON_IMPL_CUDA)
#elif defined(GGML_COMMON_IMPL_CUDA) || defined(GGML_COMMON_IMPL_HIP)
#include <cstdint>
#define GGML_TABLE_BEGIN(type, name, size) static const __device__ type name[size] = {
@ -1423,3 +1827,4 @@ GGML_TABLE_END()
#endif
#endif // GGML_COMMON_IMPL
#endif // GGML_COMMON_IMPL

244
ggml-cuda.cu
View File

@ -2,7 +2,13 @@
#include "ggml.h"
#include "ggml-backend-impl.h"
#if defined(GGML_USE_HIPBLAS)
#define GGML_COMMON_DECL_HIP
#define GGML_COMMON_IMPL_HIP
#else
#define GGML_COMMON_DECL_CUDA
#define GGML_COMMON_IMPL_CUDA
#endif
#include "ggml-common.h"
#include <algorithm>
@ -359,66 +365,6 @@ typedef void (*ggml_cuda_op_flatten_t)(
const ggml_tensor * src0, const ggml_tensor * src1, ggml_tensor * dst,
const float * src0_dd, const float * src1_dd, float * dst_dd, cudaStream_t main_stream);
// QK = number of values after dequantization
// QR = QK / number of values before dequantization
// QI = number of 32 bit integers before dequantization
#define QK4_0 32
#define QR4_0 2
#define QI4_0 (QK4_0 / (4 * QR4_0))
typedef struct {
half d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
#define QK4_1 32
#define QR4_1 2
#define QI4_1 (QK4_1 / (4 * QR4_1))
typedef struct {
half2 dm; // dm.x = delta, dm.y = min
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
#define QK5_0 32
#define QR5_0 2
#define QI5_0 (QK5_0 / (4 * QR5_0))
typedef struct {
half d; // delta
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_0 / 2]; // nibbles / quants
} block_q5_0;
static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
#define QK5_1 32
#define QR5_1 2
#define QI5_1 (QK5_1 / (4 * QR5_1))
typedef struct {
half2 dm; // dm.x = delta, dm.y = min
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_1 / 2]; // nibbles / quants
} block_q5_1;
static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
#define QK8_0 32
#define QR8_0 1
#define QI8_0 (QK8_0 / (4 * QR8_0))
typedef struct {
half d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
#define QK8_1 32
#define QR8_1 1
#define QI8_1 (QK8_1 / (4 * QR8_1))
typedef struct {
half2 ds; // ds.x = delta, ds.y = sum
int8_t qs[QK8_0]; // quants
} block_q8_1;
static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_fp16_t) + QK8_0, "wrong q8_1 block size/padding");
typedef float (*vec_dot_q_cuda_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
typedef void (*allocate_tiles_cuda_t)(int ** x_ql, half2 ** x_dm, int ** x_qh, int ** x_sc);
typedef void (*load_tiles_cuda_t)(
@ -428,174 +374,6 @@ typedef float (*vec_dot_q_mul_mat_cuda_t)(
const int * __restrict__ x_ql, const half2 * __restrict__ x_dm, const int * __restrict__ x_qh, const int * __restrict__ x_sc,
const int * __restrict__ y_qs, const half2 * __restrict__ y_ms, const int & i, const int & j, const int & k);
//================================= k-quants
#ifdef GGML_QKK_64
#define QK_K 64
#define K_SCALE_SIZE 4
#else
#define QK_K 256
#define K_SCALE_SIZE 12
#endif
#define QR2_K 4
#define QI2_K (QK_K / (4*QR2_K))
typedef struct {
uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
uint8_t qs[QK_K/4]; // quants
half2 dm; // super-block scale for quantized scales/mins
} block_q2_K;
static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
#define QR3_K 4
#define QI3_K (QK_K / (4*QR3_K))
typedef struct {
uint8_t hmask[QK_K/8]; // quants - high bit
uint8_t qs[QK_K/4]; // quants - low 2 bits
#ifdef GGML_QKK_64
uint8_t scales[2]; // scales, quantized with 8 bits
#else
uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
#endif
half d; // super-block scale
} block_q3_K;
//static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + K_SCALE_SIZE, "wrong q3_K block size/padding");
#define QR4_K 2
#define QI4_K (QK_K / (4*QR4_K))
#ifdef GGML_QKK_64
typedef struct {
half dm[2]; // super-block scales/mins
uint8_t scales[2]; // 4-bit block scales/mins
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == sizeof(half2) + QK_K/2 + 2, "wrong q4_K block size/padding");
#else
typedef struct {
half2 dm; // super-block scale for quantized scales/mins
uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_K block size/padding");
#endif
#define QR5_K 2
#define QI5_K (QK_K / (4*QR5_K))
#ifdef GGML_QKK_64
typedef struct {
half d; // super-block scale
int8_t scales[QK_K/16]; // block scales
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
#else
typedef struct {
half2 dm; // super-block scale for quantized scales/mins
uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
#endif
#define QR6_K 2
#define QI6_K (QK_K / (4*QR6_K))
typedef struct {
uint8_t ql[QK_K/2]; // quants, lower 4 bits
uint8_t qh[QK_K/4]; // quants, upper 2 bits
int8_t scales[QK_K/16]; // scales
half d; // delta
} block_q6_K;
static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_K block size/padding");
#define QR2_XXS 8
#define QI2_XXS (QK_K / (4*QR2_XXS))
typedef struct {
half d;
uint16_t qs[QK_K/8];
} block_iq2_xxs;
static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
#define QR2_XS 8
#define QI2_XS (QK_K / (4*QR2_XS))
typedef struct {
half d;
uint16_t qs[QK_K/8];
uint8_t scales[QK_K/32];
} block_iq2_xs;
static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
// 2.5625 bpw quants
#define QR2_S 8
#define QI2_S (QK_K / (4*QR2_S))
typedef struct {
half d;
uint8_t qs[QK_K/4];
uint8_t qh[QK_K/32];
uint8_t scales[QK_K/32];
} block_iq2_s;
static_assert(sizeof(block_iq2_s) == sizeof(ggml_fp16_t) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
#define QR3_XXS 8
#define QI3_XXS (QK_K / (4*QR3_XXS))
typedef struct {
half d;
uint8_t qs[3*(QK_K/8)];
} block_iq3_xxs;
static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
#define QR3_XS 8
#define QI3_XS (QK_K / (4*QR3_XS))
#if QK_K == 64
#define IQ3S_N_SCALE 2
#else
#define IQ3S_N_SCALE QK_K/64
#endif
typedef struct {
half d;
uint8_t qs[QK_K/4];
uint8_t qh[QK_K/32];
uint8_t signs[QK_K/8];
uint8_t scales[IQ3S_N_SCALE];
} block_iq3_s;
static_assert(sizeof(block_iq3_s) == sizeof(ggml_fp16_t) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
#define QR1_S 8
#define QI1_S (QK_K / (4*QR1_S))
typedef struct {
half d;
uint8_t qs[QK_K/8];
uint16_t qh[QK_K/32];
} block_iq1_s;
static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
#define QK4_NL 32
#define QR4_NL 2
#define QI4_NL (QK4_NL / (4*QR4_NL))
typedef struct {
half d;
uint8_t qs[QK4_NL/2];
} block_iq4_nl;
static_assert(sizeof(block_iq4_nl) == sizeof(ggml_fp16_t) + QK4_NL/2, "wrong iq4_nl block size/padding");
#if QK_K == 64
#define block_iq4_xs block_iq4_nl
#define QR4_XS QR4_NL
#define QI4_XS QI4_NL
#else
// QR4_XS = 8 is very slightly faster than QR4_XS = 4
#define QR4_XS 8
#define QI4_XS (QK_K / (4*QR4_XS))
typedef struct {
half d;
uint16_t scales_h;
uint8_t scales_l[QK_K/64];
uint8_t qs[QK_K/2];
} block_iq4_xs;
static_assert(sizeof(block_iq4_xs) == sizeof(ggml_fp16_t) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
#endif
#define WARP_SIZE 32
#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses
@ -3570,7 +3348,7 @@ static __device__ __forceinline__ float vec_dot_q2_K_q8_1(
#pragma unroll
for (int i = 0; i < QR2_K; ++ i) {
u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
d8[i] = __low2float(bq8_1[bq8_offset + i].ds);
}
return vec_dot_q2_K_q8_1_impl_mmvq(v, u, scales, bq2_K->dm, d8);
@ -3692,7 +3470,7 @@ static __device__ __forceinline__ float vec_dot_q3_K_q8_1(
#pragma unroll
for (int i = 0; i < QR3_K; ++i) {
u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + i].qs, iqs % QI8_1);
d8[i] = __low2half(bq8_1[bq8_offset + i].ds);
d8[i] = __low2float(bq8_1[bq8_offset + i].ds);
}
return vec_dot_q3_K_q8_1_impl_mmvq(vl, vh, u, bq3_K->scales, scale_offset, d, d8);
@ -3861,7 +3639,7 @@ static __device__ __forceinline__ float vec_dot_q4_K_q8_1(
for (int i = 0; i < QR4_K; ++i) {
const block_q8_1 * bq8i = bq8_1 + bq8_offset + i;
d8[i] = __low2half(bq8i->ds);
d8[i] = __low2float(bq8i->ds);
const int * q8 = (const int *)bq8i->qs + ((iqs/2)%4);
u[2*i+0] = q8[0];
@ -4226,7 +4004,7 @@ static __device__ __forceinline__ float vec_dot_q6_K_q8_1(
#pragma unroll
for (int i = 0; i < QR6_K; ++i) {
u[i] = get_int_from_int8_aligned(bq8_1[bq8_offset + 2*i].qs, iqs % QI8_1);
d8[i] = __low2half(bq8_1[bq8_offset + 2*i].ds);
d8[i] = __low2float(bq8_1[bq8_offset + 2*i].ds);
}
return vec_dot_q6_K_q8_1_impl_mmvq(vl, vh, u, scales, bq6_K->d, d8);
@ -4763,7 +4541,7 @@ static __device__ __forceinline__ void mul_mat_q(
*dsi_dst = *dsi_src;
} else {
float * dfi_dst = (float *) dsi_dst;
*dfi_dst = __low2half(*dsi_src);
*dfi_dst = __low2float(*dsi_src);
}
}

2
ggml-metal.m
View File

@ -336,7 +336,7 @@ static struct ggml_metal_context * ggml_metal_init(int n_cb) {
NSMutableDictionary * prep = [NSMutableDictionary dictionary];
#ifdef GGML_QKK_64
prep[@"QK_K"] = @(64);
prep[@"GGML_QKK_64"] = @(1);
#endif
MTLCompileOptions* options = [MTLCompileOptions new];

180
ggml-metal.metal
View File

@ -1,3 +1,7 @@
#define GGML_COMMON_DECL_METAL
#define GGML_COMMON_IMPL_METAL
#include "ggml-common.h"
#include <metal_stdlib>
#define GGML_COMMON_IMPL_METAL
@ -9,41 +13,6 @@ using namespace metal;
#define MIN(x, y) ((x) < (y) ? (x) : (y))
#define SWAP(x, y) { auto tmp = (x); (x) = (y); (y) = tmp; }
#define QK4_0 32
#define QR4_0 2
typedef struct {
half d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
#define QK4_1 32
typedef struct {
half d; // delta
half m; // min
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
#define QK5_0 32
typedef struct {
half d; // delta
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_0 / 2]; // nibbles / quants
} block_q5_0;
#define QK5_1 32
typedef struct {
half d; // delta
half m; // min
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_1 / 2]; // nibbles / quants
} block_q5_1;
#define QK8_0 32
typedef struct {
half d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
#define N_SIMDWIDTH 32 // assuming SIMD group size is 32
enum ggml_sort_order {
@ -2478,147 +2447,6 @@ kernel void kernel_concat(
}
}
//============================================ k-quants ======================================================
#ifndef QK_K
#define QK_K 256
#else
static_assert(QK_K == 256 || QK_K == 64, "QK_K must be 256 or 64");
#endif
#if QK_K == 256
#define K_SCALE_SIZE 12
#else
#define K_SCALE_SIZE 4
#endif
typedef struct {
uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
uint8_t qs[QK_K/4]; // quants
half d; // super-block scale for quantized scales
half dmin; // super-block scale for quantized mins
} block_q2_K;
// 84 bytes / block
typedef struct {
uint8_t hmask[QK_K/8]; // quants - high bit
uint8_t qs[QK_K/4]; // quants - low 2 bits
#if QK_K == 64
uint8_t scales[2];
#else
uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
#endif
half d; // super-block scale
} block_q3_K;
#if QK_K == 64
typedef struct {
half d[2]; // super-block scales/mins
uint8_t scales[2];
uint8_t qs[QK_K/2]; // 4-bit quants
} block_q4_K;
#else
typedef struct {
half d; // super-block scale for quantized scales
half dmin; // super-block scale for quantized mins
uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
#endif
#if QK_K == 64
typedef struct {
half d; // super-block scales/mins
int8_t scales[QK_K/16]; // 8-bit block scales
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
#else
typedef struct {
half d; // super-block scale for quantized scales
half dmin; // super-block scale for quantized mins
uint8_t scales[3*QK_K/64]; // scales and mins, quantized with 6 bits
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
// 176 bytes / block
#endif
typedef struct {
uint8_t ql[QK_K/2]; // quants, lower 4 bits
uint8_t qh[QK_K/4]; // quants, upper 2 bits
int8_t scales[QK_K/16]; // scales, quantized with 8 bits
half d; // super-block scale
} block_q6_K;
// 210 bytes / block
typedef struct {
half d;
uint16_t qs[QK_K/8];
} block_iq2_xxs;
// 66 bytes / block for QK_K = 256, so 2.0625 bpw
typedef struct {
half d;
uint16_t qs[QK_K/8];
uint8_t scales[QK_K/32];
} block_iq2_xs;
// 74 bytes / block for QK_K = 256, so 2.3125 bpw
// 2.5625 bpw quants
typedef struct {
half d;
uint8_t qs[QK_K/4];
uint8_t qh[QK_K/32];
uint8_t scales[QK_K/32];
} block_iq2_s;
typedef struct {
half d;
uint8_t qs[3*QK_K/8];
} block_iq3_xxs;
// 98 bytes / block for QK_K = 256, so 3.0625 bpw
// 3.4375 bpw
#if QK_K == 64
#define IQ3S_N_SCALE 2
#else
#define IQ3S_N_SCALE QK_K/64
#endif
typedef struct {
half d;
uint8_t qs[QK_K/4];
uint8_t qh[QK_K/32];
uint8_t signs[QK_K/8];
uint8_t scales[IQ3S_N_SCALE];
} block_iq3_s;
typedef struct {
half d;
uint8_t qs[QK_K/8];
uint16_t qh[QK_K/32];
} block_iq1_s;
// Non-linear quants
#define QK4_NL 32
typedef struct {
half d;
uint8_t qs[QK4_NL/2];
} block_iq4_nl;
#if QK_K == 64
#define block_iq4_xs block_iq4_nl
#else
typedef struct {
half d;
uint16_t scales_h;
uint8_t scales_l[QK_K/64];
uint8_t qs[QK_K/2];
} block_iq4_xs;
#endif
//====================================== dot products =========================
void kernel_mul_mv_q2_K_f32_impl(
device const void * src0,
device const float * src1,

78
ggml-quants.c
View File

@ -1,3 +1,6 @@
#define GGML_COMMON_IMPL_C
#include "ggml-common.h"
#include "ggml-quants.h"
#include "ggml-impl.h"
@ -951,7 +954,7 @@ void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
y[i].d = GGML_FP32_TO_FP16(d);
int sum = 0;
@ -966,7 +969,7 @@ void quantize_row_q8_1_reference(const float * restrict x, block_q8_1 * restrict
sum += y[i].qs[QK8_1/2 + j];
}
y[i].s = sum*d;
y[i].s = GGML_FP32_TO_FP16(sum*d);
}
}
@ -994,7 +997,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
y[i].d = GGML_FP32_TO_FP16(d);
int32x4_t accv = vdupq_n_s32(0);
@ -1010,7 +1013,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
accv = vaddq_s32(accv, vi);
}
y[i].s = d * vaddvq_s32(accv);
y[i].s = GGML_FP32_TO_FP16(d * vaddvq_s32(accv));
}
#elif defined(__wasm_simd128__)
for (int i = 0; i < nb; i++) {
@ -1033,7 +1036,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
y[i].d = GGML_FP32_TO_FP16(d);
v128_t accv = wasm_i32x4_splat(0);
@ -1049,10 +1052,11 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
accv = wasm_i32x4_add(accv, vi);
}
y[i].s = d * (wasm_i32x4_extract_lane(accv, 0) +
wasm_i32x4_extract_lane(accv, 1) +
wasm_i32x4_extract_lane(accv, 2) +
wasm_i32x4_extract_lane(accv, 3));
y[i].s = GGML_FP32_TO_FP16(
d * (wasm_i32x4_extract_lane(accv, 0) +
wasm_i32x4_extract_lane(accv, 1) +
wasm_i32x4_extract_lane(accv, 2) +
wasm_i32x4_extract_lane(accv, 3)));
}
#elif defined(__AVX2__) || defined(__AVX__)
for (int i = 0; i < nb; i++) {
@ -1077,7 +1081,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
// Quantize these floats
const float d = maxScalar / 127.f;
y[i].d = d;
y[i].d = GGML_FP32_TO_FP16(d);
const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
const __m256 mul = _mm256_set1_ps( id );
@ -1101,7 +1105,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
#if defined(__AVX2__)
// Compute the sum of the quants and set y[i].s
y[i].s = d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3)));
y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));
// Convert int32 to int16
i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15
@ -1131,7 +1135,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
// Compute the sum of the quants and set y[i].s
const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
y[i].s = d * hsum_i32_4(_mm_add_epi32(s0, s1));
y[i].s = GGML_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
// Convert int32 to int16
ni0 = _mm_packs_epi32( ni0, ni1 );
@ -1162,7 +1166,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
const float d = amax / ((1 << 7) - 1);
const float id = d ? 1.0f/d : 0.0f;
y[i].d = d;
y[i].d = GGML_FP32_TO_FP16(d);
vfloat32m4_t x0 = __riscv_vfmul_vf_f32m4(v_x, id, vl);
@ -1179,7 +1183,7 @@ void quantize_row_q8_1(const float * restrict x, void * restrict vy, int k) {
// set y[i].s
int sum = __riscv_vmv_x_s_i16m1_i16(vwrs);
y[i].s = sum*d;
y[i].s = GGML_FP32_TO_FP16(sum*d);
}
#else
GGML_UNUSED(nb);
@ -4019,10 +4023,10 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
const block_q8_1 * restrict b_y0 = &vy0[i];
const block_q8_1 * restrict b_y1 = &vy1[i];
float32x4_t summs_t = {GGML_FP16_TO_FP32(b_x0->m) * b_y0->s,
GGML_FP16_TO_FP32(b_x1->m) * b_y0->s,
GGML_FP16_TO_FP32(b_x0->m) * b_y1->s,
GGML_FP16_TO_FP32(b_x1->m) * b_y1->s};
float32x4_t summs_t = {GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y0->s),
GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y0->s),
GGML_FP16_TO_FP32(b_x0->m) * GGML_FP16_TO_FP32(b_y1->s),
GGML_FP16_TO_FP32(b_x1->m) * GGML_FP16_TO_FP32(b_y1->s)};
summs0 += summs_t;
const uint8x16_t m4b = vdupq_n_u8(0x0F);
@ -4087,7 +4091,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
const block_q8_1 * restrict y0 = &y[i + 0];
const block_q8_1 * restrict y1 = &y[i + 1];
summs += GGML_FP16_TO_FP32(x0->m) * y0->s + GGML_FP16_TO_FP32(x1->m) * y1->s;
summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s) + GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
const uint8x16_t m4b = vdupq_n_u8(0x0F);
@ -4110,8 +4114,8 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
const int32x4_t p_0 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_0l, v1_0l), v0_0h, v1_0h);
const int32x4_t p_1 = ggml_vdotq_s32(ggml_vdotq_s32(vdupq_n_s32(0), v0_1l, v1_1l), v0_1h, v1_1h);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*y0->d);
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*y1->d);
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(p_0), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(p_1), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
}
*s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs;
@ -4124,9 +4128,9 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
// Main loop
for (int i = 0; i < nb; ++i) {
const float d0 = GGML_FP16_TO_FP32(x[i].d);
const float d1 = y[i].d;
const float d1 = GGML_FP16_TO_FP32(y[i].d);
summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);
const __m256 d0v = _mm256_set1_ps( d0 );
const __m256 d1v = _mm256_set1_ps( d1 );
@ -4178,7 +4182,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
}
*s = sumf;
@ -4196,7 +4200,7 @@ void ggml_vec_dot_q4_1_q8_1(int n, float * restrict s, size_t bs, const void * r
sumi += (v0 * y[i].qs[j]) + (v1 * y[i].qs[j + qk/2]);
}
sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
}
*s = sumf;
@ -4532,8 +4536,8 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
const uint8x16_t m4b = vdupq_n_u8(0x0F);
summs0 += GGML_FP16_TO_FP32(x0->m) * y0->s;
summs1 += GGML_FP16_TO_FP32(x1->m) * y1->s;
summs0 += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
summs1 += GGML_FP16_TO_FP32(x1->m) * GGML_FP16_TO_FP32(y1->s);
// extract the 5th bit via lookup table ((b) << 4)
memcpy(&qh0, x0->qh, sizeof(qh0));
@ -4577,10 +4581,10 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
sumv0 = vmlaq_n_f32(sumv0, vcvtq_f32_s32(vaddq_s32(
ggml_vdotq_s32(vdupq_n_s32(0), v0_0lf, v1_0l),
ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*y0->d);
ggml_vdotq_s32(vdupq_n_s32(0), v0_0hf, v1_0h))), GGML_FP16_TO_FP32(x0->d)*GGML_FP16_TO_FP32(y0->d));
sumv1 = vmlaq_n_f32(sumv1, vcvtq_f32_s32(vaddq_s32(
ggml_vdotq_s32(vdupq_n_s32(0), v0_1lf, v1_1l),
ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*y1->d);
ggml_vdotq_s32(vdupq_n_s32(0), v0_1hf, v1_1h))), GGML_FP16_TO_FP32(x1->d)*GGML_FP16_TO_FP32(y1->d));
}
*s = vaddvq_f32(sumv0) + vaddvq_f32(sumv1) + summs0 + summs1;
@ -4597,7 +4601,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
const block_q5_1 * restrict x0 = &x[i];
const block_q8_1 * restrict y0 = &y[i];
summs += GGML_FP16_TO_FP32(x0->m) * y0->s;
summs += GGML_FP16_TO_FP32(x0->m) * GGML_FP16_TO_FP32(y0->s);
const v128_t m4b = wasm_i8x16_splat(0x0F);
@ -4644,7 +4648,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
wasm_i32x4_dot_i16x8(v0lfh, v1lh)),
wasm_i32x4_add(wasm_i32x4_dot_i16x8(v0hfl, v1hl),
wasm_i32x4_dot_i16x8(v0hfh, v1hh)))),
wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * y0->d)));
wasm_f32x4_splat(GGML_FP16_TO_FP32(x0->d) * GGML_FP16_TO_FP32(y0->d))));
}
*s = wasm_f32x4_extract_lane(sumv, 0) + wasm_f32x4_extract_lane(sumv, 1) +
@ -4659,14 +4663,14 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
for (int i = 0; i < nb; i++) {
const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d));
summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);
__m256i qx = bytes_from_nibbles_32(x[i].qs);
__m256i bxhi = bytes_from_bits_32(x[i].qh);
bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
qx = _mm256_or_si256(qx, bxhi);
const __m256 dy = _mm256_set1_ps(y[i].d);
const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[i].d));
const __m256i qy = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256 q = mul_sum_us8_pairs_float(qx, qy);
@ -4686,7 +4690,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
for (int i = 0; i < nb; i++) {
const __m256 dx = _mm256_set1_ps(GGML_FP16_TO_FP32(x[i].d));
summs += GGML_FP16_TO_FP32(x[i].m) * y[i].s;
summs += GGML_FP16_TO_FP32(x[i].m) * GGML_FP16_TO_FP32(y[i].s);
__m256i bx_0 = bytes_from_nibbles_32(x[i].qs);
const __m256i bxhi = bytes_from_bits_32(x[i].qh);
@ -4700,7 +4704,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
bxh = _mm_or_si128(bxh, bxhih);
bx_0 = MM256_SET_M128I(bxh, bxl);
const __m256 dy = _mm256_set1_ps(y[i].d);
const __m256 dy = _mm256_set1_ps(GGML_FP16_TO_FP32(y[i].d));
const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[i].qs);
const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
@ -4767,7 +4771,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
int sumi = __riscv_vmv_x_s_i32m1_i32(vs2);
sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
}
*s = sumf;
@ -4791,7 +4795,7 @@ void ggml_vec_dot_q5_1_q8_1(int n, float * restrict s, size_t bs, const void * r
sumi += (x0 * y[i].qs[j]) + (x1 * y[i].qs[j + qk/2]);
}
sumf += (GGML_FP16_TO_FP32(x[i].d)*y[i].d)*sumi + GGML_FP16_TO_FP32(x[i].m)*y[i].s;
sumf += (GGML_FP16_TO_FP32(x[i].d)*GGML_FP16_TO_FP32(y[i].d))*sumi + GGML_FP16_TO_FP32(x[i].m)*GGML_FP16_TO_FP32(y[i].s);
}
*s = sumf;

246
ggml-quants.h
View File

@ -1,248 +1,12 @@
#pragma once
#define GGML_COMMON_DECL_C
#include "ggml-common.h"
#include "ggml.h"
// GGML internal header
#include "ggml-impl.h"
#include <stdint.h>
#include <stddef.h>
#define QK4_0 32
typedef struct {
ggml_fp16_t d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
#define QK4_1 32
typedef struct {
ggml_fp16_t d; // delta
ggml_fp16_t m; // min
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == 2 * sizeof(ggml_fp16_t) + QK4_1 / 2, "wrong q4_1 block size/padding");
#define QK5_0 32
typedef struct {
ggml_fp16_t d; // delta
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_0 / 2]; // nibbles / quants
} block_q5_0;
static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
#define QK5_1 32
typedef struct {
ggml_fp16_t d; // delta
ggml_fp16_t m; // min
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_1 / 2]; // nibbles / quants
} block_q5_1;
static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
#define QK8_0 32
typedef struct {
ggml_fp16_t d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
#define QK8_1 32
typedef struct {
float d; // delta
float s; // d * sum(qs[i])
int8_t qs[QK8_1]; // quants
} block_q8_1;
static_assert(sizeof(block_q8_1) == 2*sizeof(float) + QK8_1, "wrong q8_1 block size/padding");
//
// Super-block quantization structures
//
// Super-block size
#ifdef GGML_QKK_64
#define QK_K 64
#define K_SCALE_SIZE 4
#else
#define QK_K 256
#define K_SCALE_SIZE 12
#endif
// 2-bit quantization
// weight is represented as x = a * q + b
// 16 blocks of 16 elements each
// Effectively 2.625 bits per weight
typedef struct {
uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
uint8_t qs[QK_K/4]; // quants
ggml_fp16_t d; // super-block scale for quantized scales
ggml_fp16_t dmin; // super-block scale for quantized mins
} block_q2_K;
static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
// 3-bit quantization
// weight is represented as x = a * q
// 16 blocks of 16 elements each
// Effectively 3.4375 bits per weight
#ifdef GGML_QKK_64
typedef struct {
uint8_t hmask[QK_K/8]; // quants - high bit
uint8_t qs[QK_K/4]; // quants - low 2 bits
uint8_t scales[2];
ggml_fp16_t d; // super-block scale
} block_q3_K;
static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 2, "wrong q3_K block size/padding");
#else
typedef struct {
uint8_t hmask[QK_K/8]; // quants - high bit
uint8_t qs[QK_K/4]; // quants - low 2 bits
uint8_t scales[12]; // scales, quantized with 6 bits
ggml_fp16_t d; // super-block scale
} block_q3_K;
static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + 12, "wrong q3_K block size/padding");
#endif
// 4-bit quantization
// 8 blocks of 32 elements each
// weight is represented as x = a * q + b
// Effectively 4.5 bits per weight
#ifdef GGML_QKK_64
typedef struct {
ggml_fp16_t d[2]; // super-block scales/mins
uint8_t scales[2]; // 4-bit block scales/mins
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + QK_K/2 + 2, "wrong q4_K block size/padding");
#else
typedef struct {
ggml_fp16_t d; // super-block scale for quantized scales
ggml_fp16_t dmin; // super-block scale for quantized mins
uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2, "wrong q4_K block size/padding");
#endif
// 5-bit quantization
// 8 blocks of 32 elements each
// weight is represented as x = a * q + b
// Effectively 5.5 bits per weight
#ifdef GGML_QKK_64
typedef struct {
ggml_fp16_t d; // super-block scale
int8_t scales[QK_K/16]; // 8-bit block scales
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
#else
typedef struct {
ggml_fp16_t d; // super-block scale for quantized scales
ggml_fp16_t dmin; // super-block scale for quantized mins
uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
#endif
// 6-bit quantization
// weight is represented as x = a * q
// 16 blocks of 16 elements each
// Effectively 6.5625 bits per weight
typedef struct {
uint8_t ql[QK_K/2]; // quants, lower 4 bits
uint8_t qh[QK_K/4]; // quants, upper 2 bits
int8_t scales[QK_K/16]; // scales, quantized with 8 bits
ggml_fp16_t d; // super-block scale
} block_q6_K;
static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + QK_K / 16 + 3*QK_K/4, "wrong q6_K block size/padding");
// This is only used for intermediate quantization and dot products
typedef struct {
float d; // delta
int8_t qs[QK_K]; // quants
int16_t bsums[QK_K/16]; // sum of quants in groups of 16
} block_q8_K;
static_assert(sizeof(block_q8_K) == sizeof(float) + QK_K + QK_K/16*sizeof(int16_t), "wrong q8_K block size/padding");
// (Almost) "true" 2-bit quantization.
// Due to the need to use blocks as per ggml design, it ends up using
// 2.0625 bpw because of the 16-bit scale for each block of 256.
typedef struct {
ggml_fp16_t d;
uint16_t qs[QK_K/8];
} block_iq2_xxs;
static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
// 2.3125 bpw quants
typedef struct {
ggml_fp16_t d;
uint16_t qs[QK_K/8];
uint8_t scales[QK_K/32];
} block_iq2_xs;
static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
// 2.5625 bpw quants
typedef struct {
ggml_fp16_t d;
uint8_t qs[QK_K/4];
uint8_t qh[QK_K/32];
uint8_t scales[QK_K/32];
} block_iq2_s;
static_assert(sizeof(block_iq2_s) == sizeof(ggml_fp16_t) + QK_K/4 + QK_K/16, "wrong iq2_s block size/padding");
// (Almost) "true" 3-bit quantization.
// Due to the need to use blocks as per ggml design, it ends up using
// 3.0625 bpw because of the 16-bit scale for each block of 256.
typedef struct {
ggml_fp16_t d;
uint8_t qs[3*QK_K/8];
} block_iq3_xxs;
static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
// 3.4375 bpw
#if QK_K == 64
#define IQ3S_N_SCALE 2
#else
#define IQ3S_N_SCALE QK_K/64
#endif
typedef struct {
ggml_fp16_t d;
uint8_t qs[QK_K/4];
uint8_t qh[QK_K/32];
uint8_t signs[QK_K/8];
uint8_t scales[IQ3S_N_SCALE];
} block_iq3_s;
static_assert(sizeof(block_iq3_s) == sizeof(ggml_fp16_t) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
typedef struct {
ggml_fp16_t d;
uint8_t qs[QK_K/8];
uint16_t qh[QK_K/32];
} block_iq1_s;
static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
// Non-linear quants
#define QK4_NL 32
typedef struct {
ggml_fp16_t d;
uint8_t qs[QK4_NL/2];
} block_iq4_nl;
static_assert(sizeof(block_iq4_nl) == sizeof(ggml_fp16_t) + QK4_NL/2, "wrong iq4_nl block size/padding");
#if QK_K == 64
#define block_iq4_xs block_iq4_nl
//typedef struct block_iq4_nl block_iq4_xs;
#else
typedef struct {
ggml_fp16_t d;
uint16_t scales_h;
uint8_t scales_l[QK_K/64];
uint8_t qs[QK_K/2];
} block_iq4_xs;
static_assert(sizeof(block_iq4_xs) == sizeof(ggml_fp16_t) + sizeof(uint16_t) + QK_K/64 + QK_K/2, "wrong iq4_xs block size/padding");
#endif
#ifdef __cplusplus
extern "C" {
#endif

192
ggml-sycl.cpp
View File

@ -3144,6 +3144,7 @@ namespace dpct
} // COPY from DPCT head files
#define GGML_COMMON_DECL_SYCL
#define GGML_COMMON_IMPL_SYCL
#include "ggml-common.h"
@ -3312,66 +3313,6 @@ typedef void (*ggml_sycl_op_flatten_t)(const ggml_tensor *src0,
const float *src1_dd, float *dst_dd,
const dpct::queue_ptr &main_stream);
// QK = number of values after dequantization
// QR = QK / number of values before dequantization
// QI = number of 32 bit integers before dequantization
#define QK4_0 32
#define QR4_0 2
#define QI4_0 (QK4_0 / (4 * QR4_0))
typedef struct dpct_type_block_q4_0 {
sycl::half d; // delta
uint8_t qs[QK4_0 / 2]; // nibbles / quants
} block_q4_0;
static_assert(sizeof(block_q4_0) == sizeof(ggml_fp16_t) + QK4_0 / 2, "wrong q4_0 block size/padding");
#define QK4_1 32
#define QR4_1 2
#define QI4_1 (QK4_1 / (4 * QR4_1))
typedef struct dpct_type_block_q4_1 {
sycl::half2 dm; // dm.x = delta, dm.y = min
uint8_t qs[QK4_1 / 2]; // nibbles / quants
} block_q4_1;
static_assert(sizeof(block_q4_1) == sizeof(ggml_fp16_t) * 2 + QK4_1 / 2, "wrong q4_1 block size/padding");
#define QK5_0 32
#define QR5_0 2
#define QI5_0 (QK5_0 / (4 * QR5_0))
typedef struct dpct_type_block_q5_0 {
sycl::half d; // delta
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_0 / 2]; // nibbles / quants
} block_q5_0;
static_assert(sizeof(block_q5_0) == sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_0 / 2, "wrong q5_0 block size/padding");
#define QK5_1 32
#define QR5_1 2
#define QI5_1 (QK5_1 / (4 * QR5_1))
typedef struct dpct_type_block_q5_1 {
sycl::half2 dm; // dm.x = delta, dm.y = min
uint8_t qh[4]; // 5-th bit of quants
uint8_t qs[QK5_1 / 2]; // nibbles / quants
} block_q5_1;
static_assert(sizeof(block_q5_1) == 2 * sizeof(ggml_fp16_t) + sizeof(uint32_t) + QK5_1 / 2, "wrong q5_1 block size/padding");
#define QK8_0 32
#define QR8_0 1
#define QI8_0 (QK8_0 / (4 * QR8_0))
typedef struct dpct_type_block_q8_0 {
sycl::half d; // delta
int8_t qs[QK8_0]; // quants
} block_q8_0;
static_assert(sizeof(block_q8_0) == sizeof(ggml_fp16_t) + QK8_0, "wrong q8_0 block size/padding");
#define QK8_1 32
#define QR8_1 1
#define QI8_1 (QK8_1 / (4 * QR8_1))
typedef struct dpct_type_block_q8_1 {
sycl::half2 ds; // ds.x = delta, ds.y = sum
int8_t qs[QK8_0]; // quants
} block_q8_1;
static_assert(sizeof(block_q8_1) == 2*sizeof(ggml_fp16_t) + QK8_0, "wrong q8_1 block size/padding");
typedef float (*vec_dot_q_sycl_t)(const void * __restrict__ vbq, const block_q8_1 * __restrict__ bq8_1, const int & iqs);
typedef void (*allocate_tiles_sycl_t)(int **x_ql, sycl::half2 **x_dm,
int **x_qh, int **x_sc);
@ -3388,137 +3329,6 @@ typedef float (*vec_dot_q_mul_mat_sycl_t)(
const int *__restrict__ y_qs, const sycl::half2 *__restrict__ y_ms,
const int &i, const int &j, const int &k);
//================================= k-quants
#ifdef GGML_QKK_64
#define QK_K 64
#define K_SCALE_SIZE 4
#else
#define QK_K 256
#define K_SCALE_SIZE 12
#endif
#define QR2_K 4
#define QI2_K (QK_K / (4*QR2_K))
typedef struct dpct_type_block_q2_K {
uint8_t scales[QK_K/16]; // scales and mins, quantized with 4 bits
uint8_t qs[QK_K/4]; // quants
sycl::half2 dm; // super-block scale for quantized scales/mins
} block_q2_K;
static_assert(sizeof(block_q2_K) == 2*sizeof(ggml_fp16_t) + QK_K/16 + QK_K/4, "wrong q2_K block size/padding");
#define QR3_K 4
#define QI3_K (QK_K / (4*QR3_K))
typedef struct dpct_type_block_q3_K {
uint8_t hmask[QK_K/8]; // quants - high bit
uint8_t qs[QK_K/4]; // quants - low 2 bits
#ifdef GGML_QKK_64
uint8_t scales[2]; // scales, quantized with 8 bits
#else
uint8_t scales[K_SCALE_SIZE]; // scales, quantized with 6 bits
#endif
sycl::half d; // super-block scale
} block_q3_K;
//static_assert(sizeof(block_q3_K) == sizeof(ggml_fp16_t) + QK_K / 4 + QK_K / 8 + K_SCALE_SIZE, "wrong q3_K block size/padding");
#define QR4_K 2
#define QI4_K (QK_K / (4*QR4_K))
#ifdef GGML_QKK_64
typedef struct {
sycl::half dm[2]; // super-block scales/mins
uint8_t scales[2]; // 4-bit block scales/mins
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == sizeof(sycl::half2) + QK_K/2 + 2, "wrong q4_K block size/padding");
#else
typedef struct dpct_type_block_q4_K {
sycl::half2 dm; // super-block scale for quantized scales/mins
uint8_t scales[3*QK_K/64]; // scales, quantized with 6 bits
uint8_t qs[QK_K/2]; // 4--bit quants
} block_q4_K;
static_assert(sizeof(block_q4_K) == 2*sizeof(ggml_fp16_t) + 3*QK_K/64 + QK_K/2, "wrong q4_K block size/padding");
#endif
#define QR5_K 2
#define QI5_K (QK_K / (4*QR5_K))
#ifdef GGML_QKK_64
typedef struct {
sycl::half d; // super-block scale
int8_t scales[QK_K/16]; // block scales
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == sizeof(ggml_fp16_t) + QK_K/2 + QK_K/8 + QK_K/16, "wrong q5_K block size/padding");
#else
typedef struct dpct_type_block_q5_K {
sycl::half2 dm; // super-block scale for quantized scales/mins
uint8_t scales[K_SCALE_SIZE]; // scales and mins, quantized with 6 bits
uint8_t qh[QK_K/8]; // quants, high bit
uint8_t qs[QK_K/2]; // quants, low 4 bits
} block_q5_K;
static_assert(sizeof(block_q5_K) == 2*sizeof(ggml_fp16_t) + K_SCALE_SIZE + QK_K/2 + QK_K/8, "wrong q5_K block size/padding");
#endif
#define QR6_K 2
#define QI6_K (QK_K / (4*QR6_K))
typedef struct dpct_type_block_q6_K {
uint8_t ql[QK_K/2]; // quants, lower 4 bits
uint8_t qh[QK_K/4]; // quants, upper 2 bits
int8_t scales[QK_K/16]; // scales
sycl::half d; // delta
} block_q6_K;
static_assert(sizeof(block_q6_K) == sizeof(ggml_fp16_t) + 13*QK_K/16, "wrong q6_K block size/padding");
#define QR2_XXS 8
#define QI2_XXS (QK_K / (4*QR2_XXS))
typedef struct dpct_type_block_iq2_xxs {
sycl::half d;
uint16_t qs[QK_K/8];
} block_iq2_xxs;
static_assert(sizeof(block_iq2_xxs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t), "wrong iq2_xxs block size/padding");
#define QR2_XS 8
#define QI2_XS (QK_K / (4*QR2_XS))
typedef struct dpct_type_block_iq2_xs {
sycl::half d;
uint16_t qs[QK_K/8];
uint8_t scales[QK_K/32];
} block_iq2_xs;
static_assert(sizeof(block_iq2_xs) == sizeof(ggml_fp16_t) + QK_K/8*sizeof(uint16_t) + QK_K/32, "wrong iq2_xs block size/padding");
#define QR3_XXS 8
#define QI3_XXS (QK_K / (4*QR3_XXS))
typedef struct dpct_type_block_iq3_xxs {
sycl::half d;
uint8_t qs[3*(QK_K/8)];
} block_iq3_xxs;
static_assert(sizeof(block_iq3_xxs) == sizeof(ggml_fp16_t) + 3*(QK_K/8), "wrong iq3_xxs block size/padding");
#define QR3_XS 8
#define QI3_XS (QK_K / (4*QR3_XS))
#if QK_K == 64
#define IQ3S_N_SCALE 2
#else
#define IQ3S_N_SCALE QK_K/64
#endif
typedef struct {
sycl::half d;
uint8_t qs[QK_K/4];
uint8_t qh[QK_K/32];
uint8_t signs[QK_K/8];
uint8_t scales[IQ3S_N_SCALE];
} block_iq3_s;
static_assert(sizeof(block_iq3_s) == sizeof(ggml_fp16_t) + 13*(QK_K/32) + IQ3S_N_SCALE, "wrong iq3_s block size/padding");
#define QR1_S 8
#define QI1_S (QK_K / (4*QR1_S))
typedef struct {
sycl::half d;
uint8_t qs[QK_K/8];
uint16_t qh[QK_K/32];
} block_iq1_s;
static_assert(sizeof(block_iq1_s) == sizeof(ggml_fp16_t) + QK_K/8 + QK_K/16, "wrong iq1_s block size/padding");
#define WARP_SIZE 32
#define MATRIX_ROW_PADDING 512 // last row of quant. matrices is a multiple of this to avoid out-of-bounds memory accesses