|
ember.cpp 0.1.0
C++ inference engine for Llama-family and paperkiln models (CPU, CUDA, WebAssembly)
|
5-bit K-quantized block structure (ggml's block_q5_K) More...
#include <quantization.h>

Public Attributes | |
| uint16_t | d |
| uint16_t | dmin |
| uint8_t | scales [12] |
| uint8_t | qh [GGML_QK_K/8] |
| uint8_t | qs [GGML_QK_K/2] |
5-bit K-quantized block structure (ggml's block_q5_K)
Q4_K's super-block scale, minimum and packed 6-bit sub-block scales, plus one high bit per weight: element l of 32-element sub-block s is d * sc[s] * ((qs[32 * (s / 2) + l] >> 4 * (s % 2) & 15) | ((qh[l] >> s & 1) << 4)) - dmin * m[s].
Definition at line 73 of file quantization.h.
| uint16_t block_q5_K::d |
Super-block scale (f16)
Definition at line 74 of file quantization.h.
| uint16_t block_q5_K::dmin |
Super-block minimum scale (f16)
Definition at line 75 of file quantization.h.
High (fifth) bit of each weight
Definition at line 77 of file quantization.h.
Low four bits of each weight
Definition at line 78 of file quantization.h.
| uint8_t block_q5_K::scales[12] |
Sub-block scales and minimums, 6 bits each
Definition at line 76 of file quantization.h.