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

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

17
ggml/src/ggml-sycl/quants.hpp
View File

@@ -48,11 +48,11 @@ template <> struct block_q_t<GGML_TYPE_Q4_0> {
};
static constexpr std::pair<int, int> get_block_offset(const int block_index, const int /* nblocks */) {
return { block_index * (traits::qk / traits::qr), 0 };
return { block_index * (QK4_0 / QR4_0), 0 };
}
static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
return { (ncols / traits::qr * nrows) + block_index * sizeof(ggml_half), 0 };
return { (ncols / QR4_0 * nrows) + block_index * sizeof(ggml_half), 0 };
}
static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
@@ -71,14 +71,12 @@ template <> struct block_q_t<GGML_TYPE_Q4_K> {
}
static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
auto nblocks = (nrows * (ncols / traits::qk));
return { nblocks * (QK_K / 2),
auto nblocks = (nrows * (ncols / QK_K));
return { nblocks * (QK_K / 2) + (block_index * K_SCALE_SIZE),
(nblocks * QK_K / 2) + (nblocks * K_SCALE_SIZE) + (block_index * sizeof(ggml_half2)) };
}
static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
constexpr size_t get_total_qs_bytes(int nblocks) { return nblocks * QK_K / 2; }
};
template <> struct block_q_t<GGML_TYPE_Q6_K> {
@@ -90,22 +88,23 @@ template <> struct block_q_t<GGML_TYPE_Q6_K> {
};
static constexpr std::pair<int, int> get_block_offset(const int block_index, const int n_blocks) {
auto low_bits_index = block_index * (traits::qk / traits::qr);
auto low_bits_index = block_index * (QK_K / QR6_K);
// the index of high bits it's after all low bits
auto high_bits_index = n_blocks * (QK_K / 2) + (block_index * (QK_K / 4));
return { low_bits_index, high_bits_index };
}
static constexpr std::pair<int, int> get_d_offset(int nrows, int ncols, const int block_index) {
auto nblocks = (nrows * (ncols / traits::qk));
auto nblocks = (nrows * (ncols / QK_K));
auto total_qs_bytes = nblocks * (QK_K / 2) + nblocks * (QK_K / 4);
auto block_scales = total_qs_bytes + block_index * (QK_K / 16);
auto sb_scale = total_qs_bytes + nblocks * (QK_K / 16);
auto sb_scale = total_qs_bytes + nblocks * (QK_K / 16) + block_index * sizeof(ggml_half);
return { block_scales, sb_scale };
}
static constexpr int block_to_q8_1_ratio() { return traits::qk / QK8_1; }
};
} // namespace ggml_sycl_reordered
#endif // GGML_SYCL_QUANTS_HPP