tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-18 05:56:00 -04:00
Code Issues Packages Projects Releases Wiki Activity

sycl: fixed semantics of block offset calculation (#14814)

Browse Source
This commit is contained in:
Alberto Cabrera Pérez
2025-07-24 11:09:57 +01:00
committed by GitHub
parent 86f5623d90
commit cb4a63aad6
2 changed files with 10 additions and 15 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
ggml/src/ggml-sycl/vecdotq.hpp
View File

@@ -350,11 +350,9 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q4_K> {
__dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr,
const sycl::half2 * q8_1_ds, const int & iqs) {
const int ib = ibx_offset.first / (QK_K / 2);
const uint8_t * base = static_cast<const uint8_t *>(vbq);
const uint8_t * qs = base + ibx_offset.first;
const uint8_t * scs = base + d_offset.first + ib * K_SCALE_SIZE;
const uint8_t * scs = base + d_offset.first;
const ggml_half2 * dms = reinterpret_cast<const ggml_half2 *>(base + d_offset.second);
const int bq8_offset = QR4_K * ((iqs / 2) / (QI8_1 / 2));
@@ -427,13 +425,11 @@ template <> struct reorder_vec_dot_q_sycl<GGML_TYPE_Q6_K> {
__dpct_inline__ float operator()(const void * __restrict__ vbq, const std::pair<int, int> ibx_offset,
const std::pair<int, int> d_offset, const int8_t * q8_1_quant_ptr, const sycl::half2 * q8_1_ds,
const int iqs) {
const int ib = ibx_offset.first / (QK_K / 2);
const uint8_t * base = static_cast<const uint8_t *>(vbq);
const uint8_t * ql = base + ibx_offset.first;
const uint8_t * qh = base + ibx_offset.second;
const int8_t * scales = reinterpret_cast<const int8_t *>(base + d_offset.first);
const ggml_half * d = (const ggml_half *) (base + d_offset.second) + ib;
const ggml_half * d = (const ggml_half *) (base + d_offset.second);
const int bq8_offset = 2 * QR6_K * (iqs / (QI6_K / 2)) + (iqs % (QI6_K / 2)) / (QI6_K / 4);
const int scale_offset = (QI6_K / 4) * (iqs / (QI6_K / 2)) + (iqs % (QI6_K / 2)) / (QI6_K / 8);