CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-21 12:16:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

bitnet : replace 1.58b with b1.58, as in the paper

Browse Source
This commit is contained in:
Francis Couture-Harpin
2024-06-28 20:38:12 -04:00
parent 0996149911
commit bfd2f21fb4
5 changed files with 7 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
ggml/src/ggml-common.h
View File

@@ -137,7 +137,7 @@ typedef sycl::half2 ggml_half2;
#endif // GGML_COMMON_DECL_CUDA || GGML_COMMON_DECL_HIP
// 1.625 bpw for BitNet 1.58b models
// 1.625 bpw for BitNet b1.58 models
#define QK1_3 64
typedef struct {
uint8_t q[(QK1_3 - 4*QK1_3/64)/5]; // 5 elements per byte (3^5 = 243 < 256)

2
ggml/src/ggml-quants.c
View File

@@ -3366,7 +3366,7 @@ size_t quantize_q2_2(const float * restrict src, void * restrict dst, int64_t nr
return nrow * row_size;
}
// ====================== 1.625 bpw (de)-quantization (BitNet 1.58b)
// ====================== 1.625 bpw (de)-quantization (BitNet b1.58)
void quantize_row_q1_3_reference(const float * restrict x, block_q1_3 * restrict y, int64_t k) {
assert(k % QK1_3 == 0);