tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-29 04:35:05 +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
convert-hf-to-gguf.py
View File

@ -300,7 +300,7 @@ class Model:
if self.ftype != gguf.LlamaFileType.ALL_F32 and extra_f16 and not extra_f32:
# TODO: cleaner model-specific per-tensor types
# NOTE: Q1_3 is only relevant for BitNet 1.58b
# NOTE: Q1_3 is only relevant for BitNet b1.58
if (
self.ftype == gguf.LlamaFileType.MOSTLY_Q1_3
and gguf.can_quantize_to_q1_3(data)

4
examples/quantize/quantize.cpp
View File

@ -26,8 +26,8 @@ static const std::vector<struct quant_option> QUANT_OPTIONS = {
{ "IQ2_M", LLAMA_FTYPE_MOSTLY_IQ2_M, " 2.7 bpw quantization", },
{ "IQ1_S", LLAMA_FTYPE_MOSTLY_IQ1_S, " 1.56 bpw quantization", },
{ "IQ1_M", LLAMA_FTYPE_MOSTLY_IQ1_M, " 1.75 bpw quantization", },
{ "Q1_3", LLAMA_FTYPE_MOSTLY_Q1_3, " 1.63 bpw for BitNet 1.58b", },
{ "Q2_2", LLAMA_FTYPE_MOSTLY_Q2_2, " 2.00 bpw for BitNet 1.58b", },
{ "Q1_3", LLAMA_FTYPE_MOSTLY_Q1_3, " 1.63 bpw for BitNet b1.58", },
{ "Q2_2", LLAMA_FTYPE_MOSTLY_Q2_2, " 2.00 bpw for BitNet b1.58", },
{ "Q2_K", LLAMA_FTYPE_MOSTLY_Q2_K, " 2.96G, +3.5199 ppl @ Llama-3-8B", },
{ "Q2_K_S", LLAMA_FTYPE_MOSTLY_Q2_K_S, " 2.96G, +3.1836 ppl @ Llama-3-8B", },
{ "IQ3_XXS",LLAMA_FTYPE_MOSTLY_IQ3_XXS," 3.06 bpw quantization", },

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);

4
src/llama.cpp
View File

@ -4186,8 +4186,8 @@ static std::string llama_model_ftype_name(llama_ftype ftype) {
case LLAMA_FTYPE_ALL_F32: return "all F32";
case LLAMA_FTYPE_MOSTLY_F16: return "F16";
case LLAMA_FTYPE_MOSTLY_BF16: return "BF16";
case LLAMA_FTYPE_MOSTLY_Q1_3: return "Q1_3 - 1.625 bpw for BitNet 1.58b";
case LLAMA_FTYPE_MOSTLY_Q2_2: return "Q2_2 - 2.000 bpw for BitNet 1.58b";
case LLAMA_FTYPE_MOSTLY_Q1_3: return "Q1_3 - 1.625 bpw for BitNet b1.58";
case LLAMA_FTYPE_MOSTLY_Q2_2: return "Q2_2 - 2.000 bpw for BitNet b1.58";
case LLAMA_FTYPE_MOSTLY_Q4_0: return "Q4_0";
case LLAMA_FTYPE_MOSTLY_Q4_1: return "Q4_1";
case LLAMA_FTYPE_MOSTLY_Q4_1_SOME_F16: