vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

ggml-quants : 1.625 bpw ternary packing for BitNet 1.58b

Browse source
This commit is contained in:
Francis Couture-Harpin 2024-06-19 12:21:08 -04:00
parent ac146628e4
commit bd807499f7
11 changed files with 594 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

6
gguf-py/gguf/constants.py
View file

@ -1023,6 +1023,8 @@ class GGMLQuantizationType(IntEnum):
F64 = 28
IQ1_M = 29
BF16 = 30
Q2_2 = 31
Q1_3 = 32
# TODO: add GGMLFileType from ggml_ftype in ggml.h
@ -1064,6 +1066,8 @@ class LlamaFileType(IntEnum):
MOSTLY_IQ4_XS = 30 # except 1d tensors
MOSTLY_IQ1_M = 31 # except 1d tensors
MOSTLY_BF16 = 32 # except 1d tensors
MOSTLY_Q2_2 = 33 # except 1d tensors
MOSTLY_Q1_3 = 34 # except 1d tensors
GUESSED = 1024 # not specified in the model file
@ -1137,6 +1141,8 @@ GGML_QUANT_SIZES: dict[GGMLQuantizationType, tuple[int, int]] = {
GGMLQuantizationType.F64: (1, 8),
GGMLQuantizationType.IQ1_M: (256, QK_K // 8 + QK_K // 16 + QK_K // 32),
GGMLQuantizationType.BF16: (1, 2),
GGMLQuantizationType.Q2_2: (32, 8),
GGMLQuantizationType.Q1_3: (64, 12 + 1),
}

50
gguf-py/gguf/quants.py
View file

@ -121,3 +121,53 @@ def quantize_q8_0(data: np.ndarray):
return __quantize_q8_0_lazy(data)
else:
return __quantize_q8_0_array(data)
__q1_3_block_size, __q1_3_type_size = GGML_QUANT_SIZES[GGMLQuantizationType.Q1_3]
def __quantize_q1_3_shape_change(s: tuple[int, ...]) -> tuple[int, ...]:
return (*s[:-1], s[-1] // __q1_3_block_size * __q1_3_type_size)
def __quantize_q1_3_rows(n: np.ndarray) -> np.ndarray:
shape = n.shape
assert shape[-1] % __q1_3_block_size == 0
n_blocks = n.size // __q1_3_block_size
blocks = n.reshape((n_blocks, __q1_3_block_size)).astype(np.float32, copy=False)
# assuming the weights are pre-scaled
blocks = (np.sign(blocks).astype(np.int8) + 1).view(np.uint8)
q48, rest = np.hsplit(blocks, (48,))
q12, q4 = np.hsplit(rest, (12,))
pow3 = np.array([1, 3, 9, 27])
q48 = q48.reshape((n_blocks, 12, 4))
q48 = np.sum(q48 * pow3.reshape((1, 1, 4)), axis=2, keepdims=True).reshape((n_blocks, 12))
q4 = np.sum(q4 * pow3.reshape((1, 4)), axis=1, keepdims=True)
q48 = q48 + (q12 * 81)
q = np.concatenate([q48, q4], axis=1);
q = ((q.astype(np.uint16) * 256) // 243).astype(np.uint8)
q = np.where(q != 0, q + 1, 0);
return q.reshape(__quantize_q1_3_shape_change(shape))
def __quantize_q1_3_array(n: np.ndarray) -> np.ndarray:
return __apply_over_grouped_rows(__quantize_q1_3_rows, arr=n, otype=np.uint8, oshape=__quantize_q1_3_shape_change(n.shape))
__quantize_q1_3_lazy = LazyNumpyTensor._wrap_fn(
__quantize_q1_3_array,
meta_noop=(np.uint8, __quantize_q1_3_shape_change),
)
def quantize_q1_3(data: np.ndarray):
if type(data) is LazyNumpyTensor:
return __quantize_q1_3_lazy(data)
else:
return __quantize_q1_3_array(data)