CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-07 09:57:00 +00:00
Code Issues Packages Projects Releases Wiki Activity

ggml : fix MUL_MAT_ID repack with Q8_K

Browse Source 
ggml-ci
This commit is contained in:
Georgi Gerganov
2025-03-24 13:07:10 +02:00
parent 9b169a4d4e
commit 87cd537a29
1 changed files with 34 additions and 33 deletions
Download Patch File Download Diff File Expand all files Collapse all files

67
ggml/src/ggml-cpu/ggml-cpu-aarch64.cpp
View File

@@ -811,7 +811,7 @@ static void quantize_q8_K_4x8(const float * GGML_RESTRICT x, void * GGML_RESTRIC
// i.e first four bsums from the first super block, followed by first four bsums from second super block and so on // i.e first four bsums from the first super block, followed by first four bsums from second super block and so on
for (int j = 0; j < QK_K * 4; j++) { for (int j = 0; j < QK_K * 4; j++) {
int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave; int src_offset = (j / (4 * blck_size_interleave)) * blck_size_interleave;
int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave; int src_id = (j % (4 * blck_size_interleave)) / blck_size_interleave;
src_offset += (j % blck_size_interleave); src_offset += (j % blck_size_interleave);
int index = (((j & 31) >> 3) << 2) + ((j >> 8) << 4) + ((j >> 6) & 3); int index = (((j & 31) >> 3) << 2) + ((j >> 8) << 4) + ((j >> 6) & 3);
@@ -5295,8 +5295,7 @@ template <> void gemv<block_q4_K, 8, 8>(int n, float * s, size_t bs, const void
ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); ggml_gemv_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
} }
template <> template <> void gemv<block_iq4_nl, 4, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
void gemv<block_iq4_nl, 4, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc); ggml_gemv_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
} }
@@ -5320,8 +5319,7 @@ template <> void gemm<block_q4_K, 8, 8>(int n, float * s, size_t bs, const void
ggml_gemm_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc); ggml_gemm_q4_K_8x8_q8_K(n, s, bs, vx, vy, nr, nc);
} }
template <> template <> void gemm<block_iq4_nl, 4, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
void gemm<block_iq4_nl, 4, 4>(int n, float * s, size_t bs, const void * vx, const void * vy, int nr, int nc) {
ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc); ggml_gemm_iq4_nl_4x4_q8_0(n, s, bs, vx, vy, nr, nc);
} }
@@ -5335,17 +5333,17 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override { bool work_size(int /* n_threads */, const struct ggml_tensor * op, size_t & size) override {
// not realy a GGML_TYPE_Q8_0 but same size. // not realy a GGML_TYPE_Q8_0 but same size.
switch (op->op) { switch (op->op) {
case GGML_OP_MUL_MAT: case GGML_OP_MUL_MAT:
size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1])); size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
return true; return true;
case GGML_OP_MUL_MAT_ID: case GGML_OP_MUL_MAT_ID:
size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1])); size = ggml_row_size(PARAM_TYPE, ggml_nelements(op->src[1]));
size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc. size = GGML_PAD(size, sizeof(int64_t)); // + padding for next bloc.
size += sizeof(int64_t) * (1+op->src[0]->ne[2]) * op->src[1]->ne[2]; size += sizeof(int64_t) * (1+op->src[0]->ne[2]) * op->src[1]->ne[2];
return true; return true;
default: default:
// GGML_ABORT("fatal error"); // GGML_ABORT("fatal error");
break; break;
} }
return false; return false;
} }
@@ -5399,12 +5397,13 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
const ggml_from_float_t from_float = ggml_get_type_traits_cpu(PARAM_TYPE)->from_float; const ggml_from_float_t from_float = ggml_get_type_traits_cpu(PARAM_TYPE)->from_float;
int64_t i11_processed = 0; int64_t i11_processed = 0;
if(PARAM_TYPE == GGML_TYPE_Q8_K) { if (PARAM_TYPE == GGML_TYPE_Q8_K) {
for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) { for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
quantize_mat_q8_K((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), 4, ne10, quantize_mat_q8_K((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), 4, ne10,
INTER_SIZE); INTER_SIZE);
} }
} else { } else {
GGML_ASSERT(PARAM_TYPE == GGML_TYPE_Q8_0);
for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) { for (int64_t i11 = ith * 4; i11 < ne11 - ne11 % 4; i11 += nth * 4) {
quantize_mat_q8_0((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), 4, ne10, quantize_mat_q8_0((float *) ((char *) src1->data + i11 * nb11), (void *) (wdata + i11 * nbw1), 4, ne10,
INTER_SIZE); INTER_SIZE);
@@ -5422,7 +5421,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
int64_t src0_start = (ith * ne01) / nth; int64_t src0_start = (ith * ne01) / nth;
int64_t src0_end = ((ith + 1) * ne01) / nth; int64_t src0_end = ((ith + 1) * ne01) / nth;
src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start; src0_start = (src0_start % NB_COLS) ? src0_start + NB_COLS - (src0_start % NB_COLS) : src0_start;
src0_end = (src0_end % NB_COLS) ? src0_end + NB_COLS - (src0_end % NB_COLS) : src0_end; src0_end = (src0_end % NB_COLS) ? src0_end + NB_COLS - (src0_end % NB_COLS) : src0_end;
if (src0_start >= src0_end) { if (src0_start >= src0_end) {
return; return;
} }
@@ -5452,7 +5451,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
const int ith = params->ith; const int ith = params->ith;
const int nth = params->nth; const int nth = params->nth;
const ggml_from_float_t from_float = ggml_get_type_traits_cpu(GGML_TYPE_Q8_0)->from_float; const ggml_from_float_t from_float = ggml_get_type_traits_cpu(PARAM_TYPE)->from_float;
// we don't support permuted src0 or src1 // we don't support permuted src0 or src1
GGML_ASSERT(nb00 == ggml_type_size(src0->type)); GGML_ASSERT(nb00 == ggml_type_size(src0->type));
@@ -5474,7 +5473,7 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
const int n_ids = ids->ne[0]; // n_expert_used const int n_ids = ids->ne[0]; // n_expert_used
const int n_as = ne02; // n_expert const int n_as = ne02; // n_expert
const size_t nbw1 = ggml_row_size(GGML_TYPE_Q8_0, ne10); const size_t nbw1 = ggml_row_size(PARAM_TYPE, ne10);
const size_t nbw2 = nbw1*ne11; const size_t nbw2 = nbw1*ne11;
const size_t nbw3 = nbw2*ne12; const size_t nbw3 = nbw2*ne12;
@@ -5486,12 +5485,13 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
GGML_ASSERT(params->wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as * sizeof(int64_t) + GGML_ASSERT(params->wsize >= (GGML_PAD(nbw3, sizeof(int64_t)) + n_as * sizeof(int64_t) +
n_as * ne12 * sizeof(mmid_row_mapping))); n_as * ne12 * sizeof(mmid_row_mapping)));
auto wdata = (char *) params->wdata; auto wdata = (char *) params->wdata;
auto wdata_src1_end = (char *) wdata + GGML_PAD(nbw3, sizeof(int64_t)); auto wdata_src1_end = (char *) wdata + GGML_PAD(nbw3, sizeof(int64_t));
int64_t * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as] int64_t * matrix_row_counts = (int64_t *) (wdata_src1_end); // [n_as]
struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12] struct mmid_row_mapping * matrix_rows = (struct mmid_row_mapping *) (matrix_row_counts + n_as); // [n_as][ne12]
// src1: float32 => block_q8_0 // src1: float32 => param type
for (int64_t i12 = 0; i12 < ne12; ++i12) { for (int64_t i12 = 0; i12 < ne12; ++i12) {
for (int64_t i11 = ith; i11 < ne11; i11 += nth) { for (int64_t i11 = ith; i11 < ne11; i11 += nth) {
from_float((float *)((char *) src1->data + i12 * nb12 + i11 * nb11), from_float((float *)((char *) src1->data + i12 * nb12 + i11 * nb11),
@@ -5537,21 +5537,22 @@ template <typename BLOC_TYPE, int64_t INTER_SIZE, int64_t NB_COLS, ggml_type PAR
int64_t src0_cur_start = (ith * ne01) / nth; int64_t src0_cur_start = (ith * ne01) / nth;
int64_t src0_cur_end = ((ith + 1) * ne01) / nth; int64_t src0_cur_end = ((ith + 1) * ne01) / nth;
src0_cur_start =
(src0_cur_start % NB_COLS) ? src0_cur_start + NB_COLS - (src0_cur_start % NB_COLS) : src0_cur_start; src0_cur_start = (src0_cur_start % NB_COLS) ? src0_cur_start + NB_COLS - (src0_cur_start % NB_COLS) : src0_cur_start;
src0_cur_end = (src0_cur_end % NB_COLS) ? src0_cur_end + NB_COLS - (src0_cur_end % NB_COLS) : src0_cur_end; src0_cur_end = (src0_cur_end % NB_COLS) ? src0_cur_end + NB_COLS - (src0_cur_end % NB_COLS) : src0_cur_end;
if (src0_cur_start >= src0_cur_end) return; if (src0_cur_start >= src0_cur_end) return;
for (int ir1 = 0; ir1 < nr1; ir1++) { for (int ir1 = 0; ir1 < nr1; ir1++) {
struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1); struct mmid_row_mapping row_mapping = MMID_MATRIX_ROW(cur_a, ir1);
const int id = row_mapping.i1; // selected expert index
const int64_t i11 = id % ne11; const int id = row_mapping.i1; // selected expert index
const int64_t i12 = row_mapping.i2; // row index in src1
const int64_t i1 = id; // selected expert index const int64_t i11 = id % ne11;
const int64_t i2 = i12; // row const int64_t i12 = row_mapping.i2; // row index in src1
const int64_t i1 = id; // selected expert index
const int64_t i2 = i12; // row
auto src1_col = (const char *) wdata + (i11 * nbw1 + i12 * nbw2); auto src1_col = (const char *) wdata + (i11 * nbw1 + i12 * nbw2);
@@ -5578,7 +5579,7 @@ static const tensor_traits<block_q4_0, 8, 8, GGML_TYPE_Q8_0> q4_0_8x8_q8_0;
static const tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K; static const tensor_traits<block_q4_K, 8, 8, GGML_TYPE_Q8_K> q4_K_8x8_q8_K;
// instance for IQ4 // instance for IQ4
static const tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_IQ4_NL> iq4_nl_4x4_q8_0; static const tensor_traits<block_iq4_nl, 4, 4, GGML_TYPE_Q8_0> iq4_nl_4x4_q8_0;
} // namespace ggml::cpu::aarch64 } // namespace ggml::cpu::aarch64