mirror of
https://github.com/kvcache-ai/ktransformers.git
synced 2026-05-05 23:50:14 +00:00
Spelling
This commit is contained in:
Iwan Kawrakow
parent
c46b0c59d0
commit
99a247e167
1 changed files with 12 additions and 12 deletions
24
third_party/llamafile/iqk_mul_mat.inc
vendored
24
third_party/llamafile/iqk_mul_mat.inc
vendored
|
|
@ -1225,7 +1225,7 @@ struct BaseDequantizer {
|
|||
float d;
|
||||
};
|
||||
|
||||
// Coped from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L1698
|
||||
// Copied from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L1698
|
||||
// MIT licensed, Copyright (c) 2024-2025 Iwan Kawrakow
|
||||
__m128i inline load_iq4nl_values_128() {
|
||||
static const uint8_t kvalues_iq4nl[16] = {1, 24, 45, 63, 79, 93, 106, 118, 129, 141, 153, 166, 181, 197, 217, 241};
|
||||
|
|
@ -1309,7 +1309,7 @@ struct DequantizerQ4K final : public BaseDequantizer<block_q4_K> {
|
|||
moonll DequantizerIQ4XS
|
||||
*/
|
||||
|
||||
// Coped from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L1775
|
||||
// Copied from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L1775
|
||||
// MIT licensed, Copyright (c) 2024-2025 Iwan Kawrakow
|
||||
__m512i inline load_iq4nl_values_512() {
|
||||
auto val256 = load_iq4nl_values_256();
|
||||
|
|
@ -1320,7 +1320,7 @@ __m512i inline load_iq4nl_values_512() {
|
|||
// Copied from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L1781
|
||||
// MIT licensed, Copyright (c) 2024-2025 Iwan Kawrakow
|
||||
struct DequantizerIQ4XS final : public BaseDequantizer<block_iq4_xs> {
|
||||
// Coped from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L1782
|
||||
// Copied from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L1782
|
||||
DequantizerIQ4XS(const void * vx, size_t bx) : BaseDequantizer(vx, bx), values(load_iq4nl_values_512()) {}
|
||||
template <typename Q8>
|
||||
inline void new_block(int i, const Q8& q8, __m256 * accd, __m512i * scales) {
|
||||
|
|
@ -1535,7 +1535,7 @@ static void mul_mat_qX_K_q8_K_T(int n, const void * vx, size_t bx, const DataInf
|
|||
|
||||
}
|
||||
}
|
||||
// Coped from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L2408
|
||||
// Copied from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L2408
|
||||
// MIT licensed, Copyright (c) 2024-2025 Iwan Kawrakow
|
||||
template <typename Q8>
|
||||
inline void compute_block(int iy, int i, float d, const Q8& q8, const __m512i * values, const __m512i * scales, __m512 * accd) {
|
||||
|
|
@ -2096,7 +2096,7 @@ struct ScaleHelperQ_0 {
|
|||
template <typename Q> inline float prepare1(const Q * y) const { return GGML_FP16_TO_FP32(y->d); }
|
||||
template <typename Q> inline float prepare1(float d, const Q * y) const { return d*prepare1(y); }
|
||||
};
|
||||
// Coped from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L8187
|
||||
// Copied from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L8187
|
||||
// MIT licensed, Copyright (c) 2024-2025 Iwan Kawrakow
|
||||
template <int min_value>
|
||||
struct ScaleHelperQ_0_1 {
|
||||
|
|
@ -2288,7 +2288,7 @@ struct Q8_0_1_Dequantizer {
|
|||
return _mm256_add_epi8(_mm256_set1_epi8(127), _mm256_loadu_si256((const __m256i *)x->qs));
|
||||
}
|
||||
};
|
||||
// end coped from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L8455
|
||||
// end copied from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L8455
|
||||
|
||||
struct Q4_0_Dequantizer {
|
||||
Dequantizer4bit b4;
|
||||
|
|
@ -2375,7 +2375,7 @@ struct Q8_0_Unpacker final : public Q_Unpacker<block_q8_0, ScaleHelperQ_0, Q8_0_
|
|||
Q8_0_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {}
|
||||
inline static int block_size() { return QK4_0; }
|
||||
};
|
||||
// Coped from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L8574
|
||||
// Copied from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L8574
|
||||
// MIT licensed, Copyright (c) 2024-2025 Iwan Kawrakow
|
||||
struct Q8_0_1_Unpacker final : public Q_Unpacker<block_q8_0, ScaleHelperQ_0_1<127>, Q8_0_1_Dequantizer> {
|
||||
Q8_0_1_Unpacker(const void * vx, size_t bx) : Q_Unpacker(vx, bx) {}
|
||||
|
|
@ -2436,7 +2436,7 @@ struct SimpleBits {
|
|||
#define HAVE_AVX512_POPCNT 0
|
||||
#endif
|
||||
|
||||
// Coped from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L7736
|
||||
// Copied from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L7736
|
||||
// with the addition of a branch that handles a missing _mm256_popcnt_epi32 instruction
|
||||
// MIT licensed, Copyright (c) 2024-2025 Iwan Kawrakow
|
||||
struct EvenSignHelper {
|
||||
|
|
@ -2494,7 +2494,7 @@ get_scale_shuffle_16
|
|||
set_scales_16
|
||||
*/
|
||||
|
||||
// Coped from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L1578
|
||||
// Copied from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L1578
|
||||
// MIT licensed, Copyright (c) 2024-2025 Iwan Kawrakow
|
||||
inline __m256i get_scale_shuffle_8(int i) {
|
||||
return _mm256_set1_epi16((2*i) | ((2*i+1) << 8));
|
||||
|
|
@ -2640,7 +2640,7 @@ inline void set_scales_16_iq(const __m256i& all_scales, __m256i * scales) {
|
|||
}
|
||||
// end copied from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L7278
|
||||
|
||||
// Coped from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L7299
|
||||
// Copied from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L7299
|
||||
// MIT licensed, Copyright (c) 2024-2025 Iwan Kawrakow
|
||||
template <typename Dequantizer>
|
||||
static void mul_mat_qX_K_q8_K_IQ_1(int n, const void * vx, size_t bx, const DataInfo& info, int nrc_x) {
|
||||
|
|
@ -2737,7 +2737,7 @@ static void mul_mat_qX_K_q8_K_IQ(int n, const void * vx, size_t bx, const DataIn
|
|||
mul_mat_qX_K_q8_K_IQ_N<Dequantizer, nrc_y>(n, vx, bx, info, nrc_x);
|
||||
#endif
|
||||
}
|
||||
// end coped from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L7299
|
||||
// end copied from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L7299
|
||||
|
||||
/*
|
||||
moonll iq1s
|
||||
|
|
@ -2828,7 +2828,7 @@ DequantizerIQ2XXS
|
|||
DequantizerIQ2XXS is important Dequantizer for DequantizerIQ1_S
|
||||
*/
|
||||
|
||||
// Coped from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L8035
|
||||
// Copied from https://github.com/ikawrakow/ik_llama.cpp/blob/474435f58b6a26bc549589966482207fee94aa60/ggml/src/iqk/iqk_mul_mat.cpp#L8035
|
||||
// MIT licensed, Copyright (c) 2024-2025 Iwan Kawrakow
|
||||
struct DequantizerIQ2XXS final : public BaseDequantizer<block_iq2_xxs> {
|
||||
DequantizerIQ2XXS(const void * vx, size_t bx) : BaseDequantizer(vx, bx) {}
|
||||
|
|
|
|||
Loading…
Add table
Add a link
Reference in a new issue