llama.cpp/gguf-py/gguf/quants.py

1270 lines
59 KiB
Python
Raw Normal View History

from __future__ import annotations
from abc import ABC, abstractmethod
from typing import Any, Callable, Sequence
from math import log2, ceil
from numpy.typing import DTypeLike
from .constants import GGML_QUANT_SIZES, GGMLQuantizationType, QK_K
from .lazy import LazyNumpyTensor
import numpy as np
def quant_shape_to_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType) -> tuple[int, ...]:
block_size, type_size = GGML_QUANT_SIZES[quant_type]
if shape[-1] % block_size != 0:
raise ValueError(f"Quantized tensor row size ({shape[-1]}) is not a multiple of {quant_type.name} block size ({block_size})")
return (*shape[:-1], shape[-1] // block_size * type_size)
def quant_shape_from_byte_shape(shape: Sequence[int], quant_type: GGMLQuantizationType) -> tuple[int, ...]:
block_size, type_size = GGML_QUANT_SIZES[quant_type]
if shape[-1] % type_size != 0:
raise ValueError(f"Quantized tensor bytes per row ({shape[-1]}) is not a multiple of {quant_type.name} type size ({type_size})")
return (*shape[:-1], shape[-1] // type_size * block_size)
# This is faster than np.vectorize and np.apply_along_axis because it works on more than one row at a time
def _apply_over_grouped_rows(func: Callable[[np.ndarray], np.ndarray], arr: np.ndarray, otype: DTypeLike, oshape: tuple[int, ...]) -> np.ndarray:
rows = arr.reshape((-1, arr.shape[-1]))
osize = 1
for dim in oshape:
osize *= dim
out = np.empty(shape=osize, dtype=otype)
# compute over groups of 16 rows (arbitrary, but seems good for performance)
n_groups = (rows.shape[0] // 16) or 1
np.concatenate([func(group).ravel() for group in np.array_split(rows, n_groups)], axis=0, out=out)
return out.reshape(oshape)
# round away from zero
# ref: https://stackoverflow.com/a/59143326/22827863
def np_roundf(n: np.ndarray) -> np.ndarray:
a = abs(n)
floored = np.floor(a)
b = floored + np.floor(2 * (a - floored))
return np.sign(n) * b
class QuantError(Exception): ...
_type_traits: dict[GGMLQuantizationType, type[__Quant]] = {}
def quantize(data: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray:
if qtype == GGMLQuantizationType.F32:
return data.astype(np.float32, copy=False)
elif qtype == GGMLQuantizationType.F16:
return data.astype(np.float16, copy=False)
elif (q := _type_traits.get(qtype)) is not None:
return q.quantize(data)
else:
raise NotImplementedError(f"Quantization for {qtype.name} is not yet implemented")
def dequantize(data: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray:
if qtype == GGMLQuantizationType.F32:
return data.view(np.float32)
elif qtype == GGMLQuantizationType.F16:
return data.view(np.float16).astype(np.float32)
elif (q := _type_traits.get(qtype)) is not None:
return q.dequantize(data)
else:
raise NotImplementedError(f"Dequantization for {qtype.name} is not yet implemented")
class __Quant(ABC):
qtype: GGMLQuantizationType
block_size: int
type_size: int
grid: np.ndarray[Any, np.dtype[np.float32]] | None = None
grid_shape: tuple[int, int] = (0, 0)
grid_map: tuple[int | float, ...] = ()
grid_hex: bytes | None = None
def __init__(self):
return TypeError("Quant conversion classes can't have instances")
def __init_subclass__(cls, qtype: GGMLQuantizationType) -> None:
cls.qtype = qtype
cls.block_size, cls.type_size = GGML_QUANT_SIZES[qtype]
cls.__quantize_lazy = LazyNumpyTensor._wrap_fn(
cls.__quantize_array,
meta_noop=(np.uint8, cls.__shape_to_bytes)
)
cls.__dequantize_lazy = LazyNumpyTensor._wrap_fn(
cls.__dequantize_array,
meta_noop=(np.float32, cls.__shape_from_bytes)
)
assert qtype not in _type_traits
_type_traits[qtype] = cls
@classmethod
def init_grid(cls):
if cls.grid is not None or cls.grid_hex is None:
return
bits_per_elem = ceil(log2(len(cls.grid_map)))
assert bits_per_elem != 0, cls.qtype.name
elems_per_byte = 8 // bits_per_elem
grid = np.frombuffer(cls.grid_hex, dtype=np.uint8)
# decode hexadecimal chars from grid
grid = grid.reshape((-1, 2))
grid = (np.where(grid > 0x40, grid + 9, grid) & 0x0F) << np.array([4, 0], dtype=np.uint8).reshape((1, 2))
grid = grid[..., 0] | grid[..., 1]
# unpack the grid values
grid = grid.reshape((-1, 1)) >> np.array([i for i in range(0, 8, 8 // elems_per_byte)], dtype=np.uint8).reshape((1, elems_per_byte))
grid = (grid & ((1 << bits_per_elem) - 1)).reshape((-1, 1))
grid_map = np.array(cls.grid_map, dtype=np.float32).reshape((1, -1))
grid = np.take_along_axis(grid_map, grid, axis=-1)
cls.grid = grid.reshape((1, 1, *cls.grid_shape))
@classmethod
@abstractmethod
def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
raise NotImplementedError
@classmethod
@abstractmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
raise NotImplementedError
@classmethod
def quantize_rows(cls, rows: np.ndarray) -> np.ndarray:
rows = rows.astype(np.float32, copy=False)
shape = rows.shape
n_blocks = rows.size // cls.block_size
blocks = rows.reshape((n_blocks, cls.block_size))
blocks = cls.quantize_blocks(blocks)
assert blocks.dtype == np.uint8
assert blocks.shape[-1] == cls.type_size
return blocks.reshape(cls.__shape_to_bytes(shape))
@classmethod
def dequantize_rows(cls, rows: np.ndarray) -> np.ndarray:
rows = rows.view(np.uint8)
shape = rows.shape
n_blocks = rows.size // cls.type_size
blocks = rows.reshape((n_blocks, cls.type_size))
blocks = cls.dequantize_blocks(blocks)
assert blocks.dtype == np.float32
assert blocks.shape[-1] == cls.block_size
return blocks.reshape(cls.__shape_from_bytes(shape))
@classmethod
def __shape_to_bytes(cls, shape: Sequence[int]):
return quant_shape_to_byte_shape(shape, cls.qtype)
@classmethod
def __shape_from_bytes(cls, shape: Sequence[int]):
return quant_shape_from_byte_shape(shape, cls.qtype)
@classmethod
def __quantize_array(cls, array: np.ndarray) -> np.ndarray:
return _apply_over_grouped_rows(cls.quantize_rows, arr=array, otype=np.uint8, oshape=cls.__shape_to_bytes(array.shape))
@classmethod
def __dequantize_array(cls, array: np.ndarray) -> np.ndarray:
cls.init_grid()
return _apply_over_grouped_rows(cls.dequantize_rows, arr=array, otype=np.float32, oshape=cls.__shape_from_bytes(array.shape))
@classmethod
def __quantize_lazy(cls, lazy_tensor: LazyNumpyTensor, /) -> Any:
pass
@classmethod
def __dequantize_lazy(cls, lazy_tensor: LazyNumpyTensor, /) -> Any:
pass
@classmethod
def can_quantize(cls, tensor: np.ndarray | LazyNumpyTensor) -> bool:
return tensor.shape[-1] % cls.block_size == 0
@classmethod
def quantize(cls, tensor: np.ndarray | LazyNumpyTensor) -> np.ndarray:
if not cls.can_quantize(tensor):
raise QuantError(f"Can't quantize tensor with shape {tensor.shape} to {cls.qtype.name}")
if isinstance(tensor, LazyNumpyTensor):
return cls.__quantize_lazy(tensor)
else:
return cls.__quantize_array(tensor)
@classmethod
def dequantize(cls, tensor: np.ndarray | LazyNumpyTensor) -> np.ndarray:
if isinstance(tensor, LazyNumpyTensor):
return cls.__dequantize_lazy(tensor)
else:
return cls.__dequantize_array(tensor)
class BF16(__Quant, qtype=GGMLQuantizationType.BF16):
@classmethod
# same as ggml_compute_fp32_to_bf16 in ggml-impl.h
def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n = blocks.view(np.uint32)
# force nan to quiet
n = np.where((n & 0x7fffffff) > 0x7f800000, (n & np.uint32(0xffff0000)) | np.uint32(64 << 16), n)
# round to nearest even
n = (np.uint64(n) + (0x7fff + ((n >> 16) & 1))) >> 16
return n.astype(np.uint16).view(np.uint8)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
return (blocks.view(np.int16).astype(np.int32) << 16).view(np.float32)
class Q4_0(__Quant, qtype=GGMLQuantizationType.Q4_0):
@classmethod
def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
imax = abs(blocks).argmax(axis=-1, keepdims=True)
max = np.take_along_axis(blocks, imax, axis=-1)
d = max / -8
with np.errstate(divide="ignore"):
id = np.where(d == 0, 0, 1 / d)
# FIXME: Q4_0's reference rounding is cursed and depends on FMA
qs = np.trunc((np.float64(blocks) * np.float64(id)) + np.float64(8.5), dtype=np.float32).astype(np.uint8).clip(0, 15)
qs = qs.reshape((n_blocks, 2, cls.block_size // 2))
qs = qs[..., 0, :] | (qs[..., 1, :] << np.uint8(4))
d = d.astype(np.float16).view(np.uint8)
return np.concatenate([d, qs], axis=-1)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d, qs = np.hsplit(blocks, [2])
d = d.view(np.float16).astype(np.float32)
qs = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1)).astype(np.int8) - np.int8(8)
return (d * qs.astype(np.float32))
class Q4_1(__Quant, qtype=GGMLQuantizationType.Q4_1):
@classmethod
def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
max = blocks.max(axis=-1, keepdims=True)
min = blocks.min(axis=-1, keepdims=True)
d = (max - min) / 15
with np.errstate(divide="ignore"):
id = np.where(d == 0, 0, 1 / d)
qs = np.trunc((blocks - min) * id + np.float32(0.5), dtype=np.float32).astype(np.uint8).clip(0, 15)
qs = qs.reshape((n_blocks, 2, cls.block_size // 2))
qs = qs[..., 0, :] | (qs[..., 1, :] << np.uint8(4))
d = d.astype(np.float16).view(np.uint8)
m = min.astype(np.float16).view(np.uint8)
return np.concatenate([d, m, qs], axis=-1)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d, rest = np.hsplit(blocks, [2])
m, qs = np.hsplit(rest, [2])
d = d.view(np.float16).astype(np.float32)
m = m.view(np.float16).astype(np.float32)
qs = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1)).astype(np.float32)
return (d * qs) + m
class Q5_0(__Quant, qtype=GGMLQuantizationType.Q5_0):
@classmethod
def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
imax = abs(blocks).argmax(axis=-1, keepdims=True)
max = np.take_along_axis(blocks, imax, axis=-1)
d = max / -16
with np.errstate(divide="ignore"):
id = np.where(d == 0, 0, 1 / d)
# FIXME: Q5_0's reference rounding is cursed and depends on FMA
q = np.trunc((np.float64(blocks) * np.float64(id)) + np.float64(16.5), dtype=np.float32).astype(np.uint8).clip(0, 31)
qs = q.reshape((n_blocks, 2, cls.block_size // 2))
qs = (qs[..., 0, :] & np.uint8(0x0F)) | (qs[..., 1, :] << np.uint8(4))
qh = np.packbits(q.reshape((n_blocks, 1, 32)) >> np.uint8(4), axis=-1, bitorder="little").reshape(n_blocks, 4)
d = d.astype(np.float16).view(np.uint8)
return np.concatenate([d, qh, qs], axis=-1)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d, rest = np.hsplit(blocks, [2])
qh, qs = np.hsplit(rest, [4])
d = d.view(np.float16).astype(np.float32)
qh = qh.view(np.uint32)
qh = qh.reshape((n_blocks, 1)) >> np.array([i for i in range(32)], dtype=np.uint32).reshape((1, 32))
ql = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
qh = (qh & np.uint32(0x01)).astype(np.uint8)
ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1))
qs = (ql | (qh << np.uint8(4))).astype(np.int8) - np.int8(16)
return (d * qs.astype(np.float32))
class Q5_1(__Quant, qtype=GGMLQuantizationType.Q5_1):
@classmethod
def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
max = blocks.max(axis=-1, keepdims=True)
min = blocks.min(axis=-1, keepdims=True)
d = (max - min) / 31
with np.errstate(divide="ignore"):
id = np.where(d == 0, 0, 1 / d)
q = np.trunc((blocks - min) * id + np.float32(0.5), dtype=np.float32).astype(np.uint8).clip(0, 31)
qs = q.reshape((n_blocks, 2, cls.block_size // 2))
qs = (qs[..., 0, :] & np.uint8(0x0F)) | (qs[..., 1, :] << np.uint8(4))
qh = np.packbits(q.reshape((n_blocks, 1, 32)) >> np.uint8(4), axis=-1, bitorder="little").reshape(n_blocks, 4)
d = d.astype(np.float16).view(np.uint8)
m = min.astype(np.float16).view(np.uint8)
return np.concatenate([d, m, qh, qs], axis=-1)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d, rest = np.hsplit(blocks, [2])
m, rest = np.hsplit(rest, [2])
qh, qs = np.hsplit(rest, [4])
d = d.view(np.float16).astype(np.float32)
m = m.view(np.float16).astype(np.float32)
qh = qh.view(np.uint32)
qh = qh.reshape((n_blocks, 1)) >> np.array([i for i in range(32)], dtype=np.uint32).reshape((1, 32))
ql = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
qh = (qh & np.uint32(0x01)).astype(np.uint8)
ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1))
qs = (ql | (qh << np.uint8(4))).astype(np.float32)
return (d * qs) + m
class Q8_0(__Quant, qtype=GGMLQuantizationType.Q8_0):
@classmethod
# Implementation of Q8_0 with bit-exact same results as reference implementation in ggml-quants.c
def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
d = abs(blocks).max(axis=1, keepdims=True) / 127
with np.errstate(divide="ignore"):
id = np.where(d == 0, 0, 1 / d)
qs = np_roundf(blocks * id)
# (n_blocks, 2)
d = d.astype(np.float16).view(np.uint8)
# (n_blocks, block_size)
qs = qs.astype(np.int8).view(np.uint8)
return np.concatenate([d, qs], axis=1)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
d, x = np.split(blocks, [2], axis=1)
d = d.view(np.float16).astype(np.float32)
x = x.view(np.int8).astype(np.float32)
return (x * d)
class Q2_K(__Quant, qtype=GGMLQuantizationType.Q2_K):
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
scales, rest = np.hsplit(blocks, [QK_K // 16])
qs, rest = np.hsplit(rest, [QK_K // 4])
d, dmin = np.hsplit(rest, [2])
d = d.view(np.float16).astype(np.float32)
dmin = dmin.view(np.float16).astype(np.float32)
# (n_blocks, 16, 1)
dl = (d * (scales & 0xF).astype(np.float32)).reshape((n_blocks, QK_K // 16, 1))
ml = (dmin * (scales >> 4).astype(np.float32)).reshape((n_blocks, QK_K // 16, 1))
shift = np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
qs = (qs.reshape((n_blocks, -1, 1, 32)) >> shift) & np.uint8(3)
qs = qs.reshape((n_blocks, QK_K // 16, 16)).astype(np.float32)
qs = dl * qs - ml
return qs.reshape((n_blocks, -1))
class Q3_K(__Quant, qtype=GGMLQuantizationType.Q3_K):
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
hmask, rest = np.hsplit(blocks, [QK_K // 8])
qs, rest = np.hsplit(rest, [QK_K // 4])
scales, d = np.hsplit(rest, [12])
d = d.view(np.float16).astype(np.float32)
# The scales are packed at 6-bit each in this pattern:
# 0: IIIIAAAA
# 1: JJJJBBBB
# 2: KKKKCCCC
# 3: LLLLDDDD
# 4: MMMMEEEE
# 5: NNNNFFFF
# 6: OOOOGGGG
# 7: PPPPHHHH
# 8: MMIIEEAA
# 9: NNJJFFBB
# 10: OOKKGGCC
# 11: PPLLHHDD
lscales, hscales = np.hsplit(scales, [8])
lscales = lscales.reshape((n_blocks, 1, 8)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 2, 1))
lscales = lscales.reshape((n_blocks, 16))
hscales = hscales.reshape((n_blocks, 1, 4)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 4, 1))
hscales = hscales.reshape((n_blocks, 16))
scales = (lscales & np.uint8(0x0F)) | ((hscales & np.uint8(0x03)) << np.uint8(4))
scales = (scales.astype(np.int8) - np.int8(32)).astype(np.float32)
dl = (d * scales).reshape((n_blocks, 16, 1))
ql = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
qh = hmask.reshape(n_blocks, -1, 1, 32) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8, 1))
ql = ql.reshape((n_blocks, 16, QK_K // 16)) & np.uint8(3)
qh = (qh.reshape((n_blocks, 16, QK_K // 16)) & np.uint8(1))
qh = qh ^ np.uint8(1) # strangely, the offset is zero when the bitmask is 1
q = (ql.astype(np.int8) - (qh << np.uint8(2)).astype(np.int8)).astype(np.float32)
return (dl * q).reshape((n_blocks, QK_K))
class Q4_K(__Quant, qtype=GGMLQuantizationType.Q4_K):
K_SCALE_SIZE = 12
@staticmethod
def get_scale_min(scales: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
n_blocks = scales.shape[0]
scales = scales.view(np.uint8)
### Unpacking the following: ###
# 0 EEAAAAAA
# 1 FFBBBBBB
# 2 GGCCCCCC
# 3 HHDDDDDD
# 4 eeaaaaaa
# 5 ffbbbbbb
# 6 ggcccccc
# 7 hhdddddd
# 8 eeeeEEEE
# 9 ffffFFFF
# 10 ggggGGGG
# 11 hhhhHHHH
scales = scales.reshape((n_blocks, 3, 4))
d, m, m_d = np.split(scales, 3, axis=-2)
sc = np.concatenate([d & 0x3F, (m_d & 0x0F) | ((d >> 2) & 0x30)], axis=-1)
min = np.concatenate([m & 0x3F, (m_d >> 4) | ((m >> 2) & 0x30)], axis=-1)
return (sc.reshape((n_blocks, 8)), min.reshape((n_blocks, 8)))
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d, rest = np.hsplit(blocks, [2])
dmin, rest = np.hsplit(rest, [2])
scales, qs = np.hsplit(rest, [cls.K_SCALE_SIZE])
d = d.view(np.float16).astype(np.float32)
dmin = dmin.view(np.float16).astype(np.float32)
sc, m = Q4_K.get_scale_min(scales)
d = (d * sc.astype(np.float32)).reshape((n_blocks, -1, 1))
dm = (dmin * m.astype(np.float32)).reshape((n_blocks, -1, 1))
qs = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1, 32)).astype(np.float32)
return (d * qs - dm).reshape((n_blocks, QK_K))
class Q5_K(__Quant, qtype=GGMLQuantizationType.Q5_K):
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d, rest = np.hsplit(blocks, [2])
dmin, rest = np.hsplit(rest, [2])
scales, rest = np.hsplit(rest, [Q4_K.K_SCALE_SIZE])
qh, qs = np.hsplit(rest, [QK_K // 8])
d = d.view(np.float16).astype(np.float32)
dmin = dmin.view(np.float16).astype(np.float32)
sc, m = Q4_K.get_scale_min(scales)
d = (d * sc.astype(np.float32)).reshape((n_blocks, -1, 1))
dm = (dmin * m.astype(np.float32)).reshape((n_blocks, -1, 1))
ql = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
qh = qh.reshape((n_blocks, -1, 1, 32)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8, 1))
ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1, 32))
qh = (qh & np.uint8(0x01)).reshape((n_blocks, -1, 32))
q = (ql | (qh << np.uint8(4))).astype(np.float32)
return (d * q - dm).reshape((n_blocks, QK_K))
class Q6_K(__Quant, qtype=GGMLQuantizationType.Q6_K):
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
ql, rest = np.hsplit(blocks, [QK_K // 2])
qh, rest = np.hsplit(rest, [QK_K // 4])
scales, d = np.hsplit(rest, [QK_K // 16])
scales = scales.view(np.int8).astype(np.float32)
d = d.view(np.float16).astype(np.float32)
d = (d * scales).reshape((n_blocks, QK_K // 16, 1))
ql = ql.reshape((n_blocks, -1, 1, 64)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
ql = (ql & np.uint8(0x0F)).reshape((n_blocks, -1, 32))
qh = qh.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
qh = (qh & np.uint8(0x03)).reshape((n_blocks, -1, 32))
q = (ql | (qh << np.uint8(4))).astype(np.int8) - np.int8(32)
q = q.reshape((n_blocks, QK_K // 16, -1)).astype(np.float32)
return (d * q).reshape((n_blocks, QK_K))
ggml-quants : ternary packing for TriLMs and BitNet b1.58 (#8151) * ggml-quants : 1.625 bpw ternary packing for BitNet 1.58b * ggml-quants : faster 1.625 bpw AVX2 vec_dot Not using a lookup table anymore makes it match q4_0 speed. * gguf-py : fix formatting * llama : remove spaces on empty line * ggml-quants : subtract 1 when back in epi8 This makes the 1.625 bpw type go faster than q4_0. Still not the fastest. * ggml-quants : Q2_2 now faster than Q4_K on with AVX2 * ggml-quants : cleanup Q1_3 code formatting * ggml-quants : ARM NEON vec_dot for q2_2 and q1_3 * ggml-quants : use ceiling division when quantizing q1_3 * convert-hf : simplify BitNet pre-quantization This still results in the exact same tensor weights and scales, but it reveals some weirdness in the current algorithm. * convert-hf : allow converting the weird BitNet 1.3B Its FFN size is 5460 which is not convenient. The offending tensors are kept in F16, which makes the final model 5.01 bpw. * bitnet : replace 1.58b with b1.58, as in the paper * ggml-quants : fix build failure on Windows * ggml-quants : attempt to fix Arm 32-bit support * ggml : add some informative comments in q1_3 vec_dot * ggml : add TQ1_0 and TQ2_0 ternary quantization types * ggml : even faster TQ2_0 * ggml : also faster TQ1_0 Same optimization as for TQ2_0 by offsetting the sum instead of the weights. This makes TQ1_0 almost as fast as Q8_0 on AVX2. * ggml : fix build issues in certain environments * ggml : add NEON vec_dot implementation for TQ1_0 and TQ2_0 * ggml : avoid directly using vmlal_high_s8, for 32-bit ARM compat The compiler seems smart enough to use the same instruction even when using vget_high_s8 instead. * ggml : remove q1_3 and q2_2 No more 1.625 bpw and 2.000 bpw, now instead using 1.6875 bpw and 2.0625 bpw with TQ1_0 and TQ2_0, respectively. * llama : remove the separate scale tensors of BitNet b1.58 They won't be needed, since the remaining ternary quant types have built-in scales. * ggml-quants : rename fields of TQ1_0 and TQ2_0 structs for consistency * ggml-quants : allow using vdotq_s32 in TQ2_0 vec_dot Not yet tested on hardware which supports it, might not work or might not even compile. But also it might. It should make the performance better on recent ARM CPUs. * ggml-quants : remove comment about possible format change of TQ2_0 Making it slightly more convenient for AVX512 but less convenient for everything else is not worth the trouble. * gguf-py : Numpy (de)quantization for TQ1_0 and TQ2_0 * ggml-quants : use roundf instead of nearest_int for TQ1_0 and TQ2_0 This does not change anything for ternary models, since their values should never end up being in halfway cases anyway. * convert : allow direct conversion to TQ1_0 and TQ2_0 The token embeddings and output tensors are kept in F16 to allow quantizing them to Q4_K and Q6_K with llama-quantize. * llama : handle fallback for TQ1_0 and TQ2_0 with Q4_0 Q4_0 is not completely symmetric (so not lossless for ternary models), but it should be good enough. * ggml-quants : allow using ARM dot product instructions for TQ1_0 * ggml-quants : deduplicate TQ1_0 and TQ2_0 __ARM_FEATURE_DOTPROD support * ggml : remove unused ggml_mul special case It would otherwise conflict with the more general optimization coming with Mamba-2. * ggml : handle TQ1_0 and TQ2_0 in dequantization-based operators * test-backend-ops : add TQ1_0 and TQ2_0 comments for later Not yet adding uncommented, because some backends like SYCL and Metal do not properly handle unknown types in supports_op for GGML_OP_MUL_MAT. (and Metal also doesn't handle it with GGML_OP_GET_ROWS) Support for TQ1_0 and TQ2_0 for other backends than CPU will be added in follow-up pull requests.
2024-09-06 03:48:47 +02:00
class TQ1_0(__Quant, qtype=GGMLQuantizationType.TQ1_0):
@classmethod
def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d = abs(blocks).max(axis=-1, keepdims=True)
with np.errstate(divide="ignore"):
id = np.where(d == 0, 0, 1 / d)
qs = np_roundf(blocks * id)
qs = (qs.astype(np.int8) + np.int8(1)).astype(np.uint8)
qs0, qs1, qh = qs[..., :(32 * 5)], qs[..., (32 * 5):(48 * 5)], qs[..., (48 * 5):]
qs0 = qs0.reshape((n_blocks, -1, 5, 32)) * np.array([81, 27, 9, 3, 1], dtype=np.uint8).reshape((1, 1, 5, 1))
qs0 = np.sum(qs0, axis=-2).reshape((n_blocks, -1))
qs1 = qs1.reshape((n_blocks, -1, 5, 16)) * np.array([81, 27, 9, 3, 1], dtype=np.uint8).reshape((1, 1, 5, 1))
qs1 = np.sum(qs1, axis=-2).reshape((n_blocks, -1))
qh = qh.reshape((n_blocks, -1, 4, 4)) * np.array([81, 27, 9, 3], dtype=np.uint8).reshape((1, 1, 4, 1))
qh = np.sum(qh, axis=-2).reshape((n_blocks, -1))
qs = np.concatenate([qs0, qs1, qh], axis=-1)
qs = (qs.astype(np.uint16) * 256 + (243 - 1)) // 243
qs = qs.astype(np.uint8)
d = d.astype(np.float16).view(np.uint8)
return np.concatenate([qs, d], axis=-1)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
qs, rest = np.hsplit(blocks, [(QK_K - 4 * QK_K // 64) // 5])
qh, d = np.hsplit(rest, [QK_K // 64])
d = d.view(np.float16).astype(np.float32)
qs0, qs1 = qs[..., :32], qs[..., 32:]
qs0 = qs0.reshape((n_blocks, -1, 1, 32)) * np.array([1, 3, 9, 27, 81], dtype=np.uint8).reshape((1, 1, 5, 1))
qs0 = qs0.reshape((n_blocks, -1))
qs1 = qs1.reshape((n_blocks, -1, 1, 16)) * np.array([1, 3, 9, 27, 81], dtype=np.uint8).reshape((1, 1, 5, 1))
qs1 = qs1.reshape((n_blocks, -1))
qh = qh.reshape((n_blocks, -1, 1, 4)) * np.array([1, 3, 9, 27], dtype=np.uint8).reshape((1, 1, 4, 1))
qh = qh.reshape((n_blocks, -1))
qs = np.concatenate([qs0, qs1, qh], axis=-1)
qs = ((qs.astype(np.uint16) * 3) >> 8).astype(np.int8) - np.int8(1)
return (d * qs.astype(np.float32))
class TQ2_0(__Quant, qtype=GGMLQuantizationType.TQ2_0):
@classmethod
def quantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d = abs(blocks).max(axis=-1, keepdims=True)
with np.errstate(divide="ignore"):
id = np.where(d == 0, 0, 1 / d)
qs = np_roundf(blocks * id)
qs = (qs.astype(np.int8) + np.int8(1)).astype(np.uint8)
qs = qs.reshape((n_blocks, -1, 4, 32)) << np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
qs = qs[..., 0, :] | qs[..., 1, :] | qs[..., 2, :] | qs[..., 3, :]
qs = qs.reshape((n_blocks, -1))
d = d.astype(np.float16).view(np.uint8)
return np.concatenate([qs, d], axis=-1)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
qs, d = np.hsplit(blocks, [QK_K // 4])
d = d.view(np.float16).astype(np.float32)
qs = qs.reshape((n_blocks, -1, 1, 32)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4, 1))
qs = (qs & 0x03).reshape((n_blocks, -1)).astype(np.int8) - np.int8(1)
return (d * qs.astype(np.float32))
class IQ2_XXS(__Quant, qtype=GGMLQuantizationType.IQ2_XXS):
ksigns: bytes = (
b"\x00\x81\x82\x03\x84\x05\x06\x87\x88\x09\x0a\x8b\x0c\x8d\x8e\x0f"
b"\x90\x11\x12\x93\x14\x95\x96\x17\x18\x99\x9a\x1b\x9c\x1d\x1e\x9f"
b"\xa0\x21\x22\xa3\x24\xa5\xa6\x27\x28\xa9\xaa\x2b\xac\x2d\x2e\xaf"
b"\x30\xb1\xb2\x33\xb4\x35\x36\xb7\xb8\x39\x3a\xbb\x3c\xbd\xbe\x3f"
b"\xc0\x41\x42\xc3\x44\xc5\xc6\x47\x48\xc9\xca\x4b\xcc\x4d\x4e\xcf"
b"\x50\xd1\xd2\x53\xd4\x55\x56\xd7\xd8\x59\x5a\xdb\x5c\xdd\xde\x5f"
b"\x60\xe1\xe2\x63\xe4\x65\x66\xe7\xe8\x69\x6a\xeb\x6c\xed\xee\x6f"
b"\xf0\x71\x72\xf3\x74\xf5\xf6\x77\x78\xf9\xfa\x7b\xfc\x7d\x7e\xff"
)
# iq2xxs_grid, but with each byte of the original packed in 2 bits,
# by mapping 0x08 to 0, 0x19 to 1, and 0x2b to 2.
grid_shape = (256, 8)
grid_map = (0x08, 0x19, 0x2b)
grid_hex = (
b"00000200050008000a00110014002000220028002a0041004400500058006100"
b"6400800082008a00a20001010401100115014001840198010002020222028202"
b"010404041004210424044004420448046004810484049004a404000502050805"
b"200546056905800591050906100640068406a406000805080808140828084108"
b"440850085208880804094009020a140a01100410101021104010601084109010"
b"951000110811201150115a118011241245120014081420142514491480141815"
b"6215001616160118041810184018811800190519a019511a002002200a204420"
b"6120802082202921482100220222012404241024402456240025412564259026"
b"082820289428442a014004401040184021402440404048405640604081408440"
b"9040004120416141804185410142104248425642684200440844204480449944"
b"124524450046014804481048404845480049584961498249454a904a00500850"
b"1150195020508050885004514251a4519152905492540a550156545600581158"
b"195864584059085a046010604060686000615561186260620064056410651265"
b"84654268008002800a8041808280048118814081118201840484108415844084"
b"608400854685948509864086608602880489118a0490109024904090a1901691"
b"8091459200942294449451958198209902a050a085a009a100a218a450a804a9"
)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d, qs = np.hsplit(blocks, [2])
d = d.view(np.float16).astype(np.float32)
qs = qs.view(np.uint32).reshape(n_blocks, -1, 2)
db = d * (np.float32(0.5) + (qs[..., 1] >> 28).astype(np.float32)) * np.float32(0.25)
db = db.reshape((n_blocks, -1, 1, 1))
# get the sign indices and unpack the bits
signs = qs[..., 1].reshape((n_blocks, -1, 1)) >> np.array([0, 7, 14, 21], dtype=np.uint32).reshape((1, 1, 4))
ksigns = np.frombuffer(cls.ksigns, dtype=np.uint8).reshape((1, 1, 1, 128))
signs = (signs & np.uint32(0x7F)).reshape((n_blocks, -1, 4, 1))
signs = np.take_along_axis(ksigns, signs, axis=-1)
signs = signs.reshape((n_blocks, -1, 4, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 1, 8))
signs = signs & np.uint8(0x01)
signs = np.where(signs == 0, np.float32(1), np.float32(-1))
signs = signs.reshape((n_blocks, -1, 4, 8))
assert cls.grid is not None
grid = np.take_along_axis(cls.grid, qs[..., 0].copy().view(np.uint8).reshape((n_blocks, -1, 1, 1)), axis=-2)
grid = grid.reshape((n_blocks, -1, 4, 8))
return (db * grid * signs).reshape((n_blocks, -1))
class IQ2_XS(__Quant, qtype=GGMLQuantizationType.IQ2_XS):
# iq2xs_grid, but with each byte of the original packed in 2 bits,
# by mapping 0x08 to 0, 0x19 to 1, and 0x2b to 2.
grid_shape = (512, 8)
grid_map = (0x08, 0x19, 0x2b)
grid_hex = (
b"00000200050008000a0011001400160019002000220025002800410044004600"
b"49005000520055005800610064008000820085008800910094009900a0000101"
b"04010601090110011201150118011a0121012401400142014501480151015401"
b"6001680181018401900100020202050208021102140220024102440250025502"
b"80028a0201040404060409041004120415041804210424044004420445044804"
b"5104540456046004810484049004000502050505080511051405200541054405"
b"500561058005010604061006260640064206840600080208050808080a081108"
b"14082008250841084408500858088008a008aa08010904091009400981098909"
b"000a200a280a960aa00a01100410061009101010121015101810211024104010"
b"4210451048105110541060106a10811084109010001102110511081111111411"
b"2011411144115011801194119611011204120612101240126012001402140514"
b"0814111414142014411444144914501464148014011504151015401500161416"
b"49160118041810181218401854188618001905196619511aa91a002002200520"
b"08200a201120142020204120442050208020a020012104211021402148216521"
b"002222228022a82201240424102429244024002541255225992501261a26a626"
b"002808280a28202855288828a22868299029082a202a822a882a8a2a01400440"
b"0640094010401240154018402140244040404240454048404a40514054406040"
b"6540814084409040004102410541084111411441204141414441504180418541"
b"a241014204421042124229424042004402440544084411441444194420444144"
b"4444504480449444014504451045244540459a4500460a464446504601480448"
b"1048404845485448624800491149444950496949044a00500250055008501150"
b"145020502850415044505050805001510451105115514051425100524452aa52"
b"0154045410542154405460548154a154005508558055885521566856a1560058"
b"14584158505899581a5940594259855a0160046010604060546062608660a960"
b"006124624a62926200641664106540654565a46501686a682569066a546a626a"
b"00800280058008801180148020802a8041804480508080808280a880aa800181"
b"0481068110814081518159810082208280828282a082a8820184048410841284"
b"158440846084898400854485a58518866a860088088825885a8880888288a888"
b"0689228a808a888a968aa88a0190049010904090569084900091229164915692"
b"89920094059444945094589429959095929541965198a6984999159a609a00a0"
b"02a008a00aa020a02aa0a0a051a159a1a6a100a202a208a22aa280a2a0a240a4"
b"95a465a698a60aa820a822a828a8a0a8a8a804a984a986a928aa2aaa91aaaaaa"
)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d, rest = np.hsplit(blocks, [2])
qs, scales = np.hsplit(rest, [2 * QK_K // 8])
d = d.view(np.float16).astype(np.float32)
qs = qs.view(np.uint16)
scales = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
scales = (scales & 0x0F).reshape((n_blocks, -1))
db = d * (np.float32(0.5) + scales) * np.float32(0.25)
db = db.reshape((n_blocks, -1, 1, 1))
# get the sign indices and unpack the bits
signs = np.frombuffer(IQ2_XXS.ksigns, dtype=np.uint8).reshape(1, 1, 128)
signs = np.take_along_axis(signs, (qs >> 9).reshape((n_blocks, -1, 1)), axis=-1)
signs = signs.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8))
signs = signs & np.uint8(0x01)
signs = np.where(signs == 0, np.float32(1), np.float32(-1))
signs = signs.reshape((n_blocks, -1, 2, 8))
assert cls.grid is not None
grid = np.take_along_axis(cls.grid, (qs & np.uint16(511)).reshape((n_blocks, -1, 1, 1)), axis=-2)
grid = grid.reshape((n_blocks, -1, 2, 8))
return (db * grid * signs).reshape((n_blocks, -1))
class IQ2_S(__Quant, qtype=GGMLQuantizationType.IQ2_S):
# iq2s_grid, but with each byte of the original packed in 2 bits,
# by mapping 0x08 to 0, 0x19 to 1, and 0x2b to 2.
grid_shape = (1024, 8)
grid_map = (0x08, 0x19, 0x2b)
grid_hex = (
b"00000200050008000a0011001400160019002000220025002800410044004600"
b"490050005200550058006100640066006900800082008500880091009400a000"
b"a500aa0001010401060109011001120115011801210124014001420145014801"
b"510154015601590160016501680181018401900192019501a101a40100020202"
b"050208021102140220022a02410244024602490250025502800285028a029402"
b"a202010404040604090410041204150418042104240426042904400442044504"
b"48044a0451045404560459046004620465048104840486048904900495049804"
b"a104a40400050205050508050a05110514051605190520052505280541054405"
b"46054905500552055505580561056405800582058505880591059405a0050106"
b"0406060609061006150640064506480651065406600681068406900600080208"
b"050808081108140816081908200825082a084108440846084908500852085508"
b"580861086408800885089408aa08010904091009120915091809210940094509"
b"480951095409600981099009000a110a140a220a280a2a0a500a990a01100410"
b"0610091010101210151018102110241026104010421045104810511054105610"
b"59106010621065106810811084108610901095109810a110a410001102110511"
b"08110a1111111411161119112011221125112811411144114611491150115211"
b"5511581161116411801182118511881191119411011204120912101215122112"
b"2412401245125112541281128412901200140214051408141114141416141914"
b"2014251428144114441446144914501452145514581461146414801482148514"
b"881491149414a014011504150615091510151215151518152115241540154215"
b"4515481551155415601581158415901500160516081611161416201641164416"
b"50168016aa160118041806180918101815181818211840184218451848185118"
b"541860188118841800190219051908191119141920194119441950196919a219"
b"041a101a401a561a00200220052008201120142016201920202025202a204120"
b"4420502052205520642080208a209420aa200121042110211221152121214021"
b"4221452151215421602181218421902100220a22222228222a22442250228822"
b"8a22a82201240424062409241024152418242124242440244224452448245124"
b"5424602481248424902400250525082511251425202541254425502566258025"
b"0126042610264026592600280528112814284128442850288a28aa2801290429"
b"102995290a2a222a642a882a8a2a014004400640094010401240154018401a40"
b"21402440264040404240454048404a4051405440564059406040624065408140"
b"8440904095409840a140a4400041024105410841114114411641194120412241"
b"2541414144414641494150415241554158416141644180418241854188419141"
b"9441a04101420442104212421542184224424042454248425142544260428142"
b"844200440244054408440a441144144416441944204422442544284441444444"
b"46444944504452445544584461446444804482448544884491449444a0440145"
b"0445064509451045124515451845214524454045424545454845514554456045"
b"6a4581458445904500460246054608461146144620464146444650468046a546"
b"0148044809481048124815481848214824484048424845484848514854486048"
b"84489048004902490549084911491449204941494449504980499649014a044a"
b"104a404a00500250055008501150145016501950205022502550285041504450"
b"4650495050505250555058506150645080508250855088509150945001510451"
b"0651095110511251155118512151245140514251455148515151545160518151"
b"8451905100520552085211521452205241524452505269528052015404540654"
b"0954105412541554185421542454405442544554485451545454605481548454"
b"9054005502550555085511551455205541554455505580550156045610562656"
b"405600580258055808581158145820584158445850585a588058015904591059"
b"4059005a195a855aa85a01600460066010601260156018602160246040604560"
b"4860516054606060846090600061026105610861116114612061416144615061"
b"806199610462106240625662a162006405640864116414642064416444645064"
b"806401650465106540654a656865926500669466016804681068656898680069"
b"2a69426aa16a0080028005800880118014801980208025804180448050805280"
b"5580588061808080858091809480018104810981108112811581188121812481"
b"408142814581488151815481818184819081a981008205820a82118214824182"
b"4482508201840484068409841084128415841884218440844284458448845184"
b"5484608481848484908400850285058508851185148520854185448550858085"
b"8a85018604861086298640860088058811881488418844885088a28801890489"
b"40896589228a588a5a8a828aa28a019004900990109012901590189024904090"
b"4290459048905190549060908190849090900091059111911491419144915091"
b"5a910192049210924092a6920094029405940894119414942094419444945094"
b"8094969401950495109540959895a19500964696649601980498109826984098"
b"a998009949995299909a00a005a00aa014a022a02aa041a044a050a0a2a0aaa0"
b"40a165a102a20aa222a228a22aa282a288a28aa2a8a201a404a410a440a489a4"
b"a4a400a519a551a60aa828a8a2a854a986a908aa0aaa20aa22aa28aa88aaaaaa"
)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d, rest = np.hsplit(blocks, [2])
qs, rest = np.hsplit(rest, [QK_K // 8])
signs, rest = np.hsplit(rest, [QK_K // 8])
qh, scales = np.hsplit(rest, [QK_K // 32])
d = d.view(np.float16).astype(np.float32)
scales = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
scales = (scales & 0x0F).reshape((n_blocks, -1))
db = d * (np.float32(0.5) + scales) * np.float32(0.25)
db = db.reshape((n_blocks, -1, 1, 1))
# unpack the sign bits
signs = signs.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8))
signs = signs & np.uint8(0x01)
signs = np.where(signs == 0, np.float32(1), np.float32(-1))
signs = signs.reshape((n_blocks, -1, 2, 8))
qh = qh.reshape((n_blocks, -1, 1)) >> np.array([0, 2, 4, 6], dtype=np.uint8).reshape((1, 1, 4))
qs = qs.astype(np.uint16) | ((qh & 0x03).astype(np.uint16) << 8).reshape((n_blocks, -1))
assert cls.grid is not None
grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
grid = grid.reshape((n_blocks, -1, 2, 8))
return (db * grid * signs).reshape((n_blocks, -1))
class IQ3_XXS(__Quant, qtype=GGMLQuantizationType.IQ3_XXS):
grid_shape = (256, 4)
grid_map = (0x04, 0x0c, 0x14, 0x1c, 0x24, 0x2c, 0x34, 0x3e)
grid_hex = (
b"0000020004001100130017002000220031004200730075000101030110011201"
b"2101250130013201410154017001000202020402110220022202310233023702"
b"5102570275020103070310031203250370031304370444045704730475040105"
b"0705320552053506640610071407160743076107011003101010121021102310"
b"3010321034104710501000110211111120112211011203121012121221123012"
b"7212001302132013311346136613011405145014201524154615711505162217"
b"4017002002201120132020202220262031204220012103210521102112212121"
b"3021632167217021002202221122172220222222372240225522012310231423"
b"7023742335245324032527254125742501270327162745270130103012302130"
b"2330503065307230003102312031313144314631013203321032253252327232"
b"1133333330344734723400350635223555351436363663363337603704401740"
b"3540374053405740744120423742404260426642074345430444514464442545"
b"4345704505471047124730471250415070500051065126515551145232527252"
b"0253535310542354275472540255315550562457425724604460466064602161"
b"6161176264623063366344640565526533660367216703700570077010703270"
b"5270267140711272457252720073157333736073217441740075027524753076"
)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d, rest = np.hsplit(blocks, [2])
qs, scales = np.hsplit(rest, [QK_K // 4])
d = d.view(np.float16).astype(np.float32)
scales = scales.view(np.uint32)
db = d * (np.float32(0.5) + (scales >> 28).astype(np.float32)) * np.float32(0.5)
db = db.reshape((n_blocks, -1, 1, 1))
# get the sign indices and unpack the bits
signs = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 7, 14, 21], dtype=np.uint32).reshape((1, 1, 4))
ksigns = np.frombuffer(IQ2_XXS.ksigns, dtype=np.uint8).reshape((1, 1, 1, 128))
signs = (signs & np.uint32(0x7F)).reshape((n_blocks, -1, 4, 1))
signs = np.take_along_axis(ksigns, signs, axis=-1)
signs = signs.reshape((n_blocks, -1, 4, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 1, 8))
signs = signs & np.uint8(0x01)
signs = np.where(signs == 0, np.float32(1), np.float32(-1))
signs = signs.reshape((n_blocks, -1, 4, 8))
assert cls.grid is not None
grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
grid = grid.reshape((n_blocks, -1, 4, 8))
return (db * grid * signs).reshape((n_blocks, -1))
class IQ3_S(__Quant, qtype=GGMLQuantizationType.IQ3_S):
grid_shape = (512, 4)
grid_map = (0x01, 0x03, 0x05, 0x07, 0x09, 0x0b, 0x0d, 0x0f)
grid_hex = (
b"0000010002000500070010001100120014001600200021002500330040004200"
b"4500470051005300600062007100740077000001010102010401100111011501"
b"2001230127013101350144016101650172010002010205020702100213021602"
b"2102250230023402420245024702510253027002730203031103150320032203"
b"3103330336034403500352036703710375030004130417042104240432044004"
b"4304510470040205040520052205260533054105450547056605730506061106"
b"1306310652067106000702070407200722072607330750075407001001100210"
b"0410101011101310151017102010221031103410361054105610611072100011"
b"0111031106111011141121113011331141115011521170117611001212121512"
b"1712201224123212401243125512601272120113041307131013131321132713"
b"3013341341136213701303140514121414143114331442144614501454140115"
b"1015131521153015321551152016241627164416461601170317101712172117"
b"3517411762177017002001200320052007201020122014201620212023202720"
b"3020322041204320452050205220672070207320752000210221102113211721"
b"2221252131213421422151210122042207222122232230223722412253225722"
b"7122742200230223052311232223242331233323422350236623012407242024"
b"2324322435244124722475240425112522253725402553257025002602260726"
b"2126552661260527112726273027432750270230113013301530173022303130"
b"3330353042304430473051306330713001310331053114312131233140316031"
b"7231763100321232203232323432503201331033143321332333273330334133"
b"4333473355337333033411341634223431345234603464340135103512352535"
b"3235443556357335163641360137033720372237353700400440124020402440"
b"2740324041405040704002410741114113412241304135414341514155410142"
b"0342104215422142334240425742624270420443114313432043224331433543"
b"0044024424443744404471440545074521456245134634466046104715473047"
b"4347514702501050145022504050445047505250665074500151035105511251"
b"2151325172510052115223523052365253520253075310532753445351536553"
b"7353015404542054325446541255265551555355425602570457225711601360"
b"1560316033606060006120612761646112623462426255626262706200631463"
b"2163406325644364626400650365346560650566406611671367007004700770"
b"2070227036704070547062700271117124714371457101720472107216722172"
b"3072517202733273357353730174057413742074507422754275027631760077"
)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d, rest = np.hsplit(blocks, [2])
qs, rest = np.hsplit(rest, [QK_K // 4])
qh, rest = np.hsplit(rest, [QK_K // 32])
signs, scales = np.hsplit(rest, [QK_K // 8])
d = d.view(np.float16).astype(np.float32)
scales = scales.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
scales = (scales & 0x0F).reshape((n_blocks, -1))
db = d * (1 + 2 * scales)
db = db.reshape((n_blocks, -1, 1, 1))
# unpack the sign bits
signs = signs.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8).reshape((1, 1, 8))
signs = signs & np.uint8(0x01)
signs = np.where(signs == 0, np.float32(1), np.float32(-1))
signs = signs.reshape((n_blocks, -1, 4, 8))
qh = qh.reshape((n_blocks, -1, 1)) >> np.array([i for i in range(8)], dtype=np.uint8)
qh = (qh & 0x01).astype(np.uint16).reshape((n_blocks, -1))
qs = qs.astype(np.uint16) | (qh << 8)
assert cls.grid is not None
grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
grid = grid.reshape((n_blocks, -1, 4, 8))
return (db * grid * signs).reshape((n_blocks, -1))
class IQ1_S(__Quant, qtype=GGMLQuantizationType.IQ1_S):
# iq1s_grid, with each byte packed into 2 bits
# -1, 0, 1 <=> 0, 1, 2
grid_shape = (2048, 8)
grid_map = (-1, 0, 1)
grid_hex = (
b"00000200050008000a00110015002000220028002a0045005100540056006500"
b"8000820088008a009500a000a200a800aa000401050111011401160119011a01"
b"2501410146014901520155015a0161016401660168018501910194019601a501"
b"0002020208020a0215022002220228022a024502510259026402690280028202"
b"88028a02910295029902a002a202a802aa021104140416042504410449045504"
b"5a046404650491049904a5040105040505050605150518051a05290540054505"
b"4a0550055105540555055605590560056205650568056a058105910595059805"
b"9a05a105a405a505a605a9051406190641064406500652065506580660066106"
b"6606690685069106940699060008020808080a0815082008220828082a084508"
b"5108560865088008820888088a089508a008a208a808aa080509110914091909"
b"2409250941095009510955096109640969099109940996099909a509000a020a"
b"080a0a0a150a200a220a280a2a0a450a510a590a610a650a800a820a850a880a"
b"8a0a950aa00aa20aa80aaa0a1010111014101910241025104110441050105510"
b"58106110641065106910911094109610a110a510011104110611091110111211"
b"1511181121112411291145114a11501151115211541155115611591160116511"
b"841192119511a111a41111121412161225124012461249125212551258125a12"
b"641266128512911294129612a512011406140914141415141814191421142614"
b"41144514461448144a1451145414551456145914621465146814841489149014"
b"94149514981499149a14a114a414a514a914021505150a151115141515151615"
b"191520152215251528152a154115441545154615511552155415551556155915"
b"5a1561156415651566156915801582158415851588158a159015911594159515"
b"961599159a15a015a215a51501160416051606161516161618161a1621162616"
b"401642164416451648164a165116551656165816591661166416651668166916"
b"6a1686168a1692169516a416a916111816182518411844184618491850185518"
b"58185a1860186118641866186918851891189418a5181019121915191a192119"
b"25194219441945194819511954195519561959195a19601965196a1989199119"
b"921995199819a119a619a919091a161a241a261a441a461a491a501a521a551a"
b"581a611a661a691a851a911a961a9a1a0020022008200a201520202022202520"
b"28202a20452051205920612065208020822088208a209520a020a220a520a820"
b"aa2005211121142119212521422144214921552158215a216121642165216621"
b"8521902196219921a521012208220a22112215222022222228222a2245225122"
b"562259226522812288228a2291229522a022a222a822aa220524142416241924"
b"252444244524462449245224552458245a2466248524912494249924a124a524"
b"0925152521252925402545254825512554255525592562256525682589259025"
b"9425952598259a25a125a425a625a92505261026122619262526412649265526"
b"6026612669268426862690269a260028022808280a2815282028222828282a28"
b"45285128542865288028822888288a28a028a228a828aa280929112914291929"
b"2529462949295229552961296429662969298529902996299929a429a529002a"
b"022a082a0a2a202a222a282a2a2a452a512a562a592a652a802a822a882a8a2a"
b"952aa02aa22aa82aaa2a054011401640254049405240554058405a4061406440"
b"664094409940a140a6400041014104410641094112411541164118411a412141"
b"26412941454148414a41514154415541564159415a41654168416a4181418441"
b"8641904192419541a041a141a241054211421442164225424142524255425a42"
b"6442694289429442a5420144154419442944454448444a445144544455445644"
b"61446244654468446a44814486448944904492449544a044a144a94401450245"
b"05450a4511451445154516451945204525452a45414544454545464549455045"
b"5145544555455645584559456145644565456645694582458445854588459145"
b"94459545964599459a45a545a845aa450146054609461446154618461a462146"
b"2446294640464246454648465046514652465546564659466246654668468146"
b"85468a4694469546a146a446a6460548114815481a4825484248494850485548"
b"5848614864486648694885489148944896489948a5480149054906490a491049"
b"144915491849214924492649404945494a495149524954495549564959496049"
b"6249654966496a49864989499249954996499849a149a449a649a949164a444a"
b"464a494a554a584a5a4a644a694a944aa54a0150045005500650095012501550"
b"1a50215024502950405045504850515054505550565059506550685086508950"
b"95509850a050a150a650a9500551085109510a51115114511551165118511951"
b"20512551265128512a5141514451455146514951505151515251545155515651"
b"585159515a51615164516551665169518251855191519451955196519951a051"
b"a551aa5101520652125215521a5221522452425245524a525152545255525652"
b"595262526552855290529252955299529a52a452045405541154145415541654"
b"185419542154255428542a54415444544554465449544a545054515454545554"
b"5654585459545a54615462546454655466546954805488548a54915494549554"
b"96549954a154a454a554aa540155025504550555065509551055115512551455"
b"1555165519551a55215524552555265529554055415542554455455546554855"
b"4955505551555255545555555655585559555a55605561556455655566556855"
b"69556a5581558455855589558a559055915594559555965598559955a155a455"
b"a555a655a9550056015602560456065608560956115614561556185619562056"
b"2156225624562556265628562956415645564656485649564a56505651565256"
b"545655565656585659565a566156645665566956825685568656885689568a56"
b"915695569a56a256a556a656a856a95604580558065809581058155818582158"
b"2a58455848584a58515854585558565858585958605862586458655882588958"
b"9058925895589858a158a9580159025905590a59115914591559165919592559"
b"41594459455946594959505951595259545955595659585959595a5961596459"
b"655966596959815985598959915994599559965998599959a559045a085a155a"
b"1a5a205a255a265a295a455a485a495a515a555a565a585a595a625a655a685a"
b"6a5a815a8a5a925a955a965a985a9a5aa15a0560146016601960256044605060"
b"5560566058605a60616064606660696081609660a56001610461066109611261"
b"15612161226126612961456149615161556156615961656166616a6184618a61"
b"92619561a161a661a96111621662196240624162466255625662586260628562"
b"91629662a56211641264156416641a6421642664296440644264456448644a64"
b"516454645564566459645a646064626465648464856489649064926494649564"
b"966498649a64a164a464a964056508650a651165156516651965446545654665"
b"496550655165546555655665596561656465656566656965866589658a659165"
b"9565966599659a65a265a565a665a86502660966156620662666286629664066"
b"456648664a66516654665566566658665a666066656668668066826685668a66"
b"9466966698669966a066a466a666aa661668196825684168526855685a686168"
b"6968856891689868a66801690469106915692169246926692969406941694569"
b"4669486951695469556956695969606965696a69826984698a699569a169a469"
b"a569a969116a166a186a416a446a496a506a556a586a5a6a646a656a696a866a"
b"946a986a9a6aa66a0080028008800a802080228028802a804580508051805480"
b"5680598065808080828088808a809580a080a280a880aa800581118114811681"
b"1981258141814481498150815281558156815881598164816681698185818981"
b"948196819981a5810082028208820a8215822082228228822a82518254825982"
b"65828082828288828a829582a082a282a882aa82148419844184448451845584"
b"5a846184648469849484998401850985128515851a8526852985408541854585"
b"4885518554855585568559855a856585668568856a8581858485868589859085"
b"928595859885a68511861686198625864186448649864a865086558659865a86"
b"618666866a86858691869a86a4860088028808880a8815882088228828882a88"
b"41884588518854885988658869888088828888888a889588a088a288a888aa88"
b"05890689118914891689258941894489468949895089528955895a8961896489"
b"858996899989a589008a028a088a0a8a158a208a228a288a2a8a458a518a548a"
b"568a808a828a888a8a8a958aa08aa28aa88aaa8a059011901690189019902590"
b"419046904990559058905a9069906a9085909190949096909990a59001910491"
b"069109911091159118911a912191249126912991409145915091519154915591"
b"569159916291659184918691929195919891a191a491a691a991059211921492"
b"19922592449246924992509252925592589266926992859294929692a9920194"
b"04940694109415941894269440944a9451945494559456945894599460946194"
b"62946594849486949294949495949894a194a9940095059508950a9510951195"
b"14951595169519952195259529952a9541954495459546954995509551955295"
b"549555955695589559955a956195649565956695699581958595889591959295"
b"94959595969599959a95a095a295a595a895aa95019604961096159619962096"
b"2696299645964896499651965296559656965996659668968296849689968a96"
b"929694969596a496a696a9960598169819982598419846985098529855985698"
b"5a98649865988598919896989998a59804990699099910991299159918991a99"
b"209921992499269940994299459948994a995199549955995699599962996599"
b"66996a99819984999099929995999a99a199a699059a159a259a449a469a499a"
b"509a559a589a619a859a919a949a959a969a00a002a008a00aa015a020a022a0"
b"28a02aa045a051a054a056a059a080a082a088a08aa095a0a0a0a2a0a8a0aaa0"
b"05a109a111a114a116a119a11aa146a149a151a155a158a15aa161a164a185a1"
b"90a192a196a199a102a208a20aa210a219a222a228a22aa245a251a256a259a2"
b"65a280a282a288a28aa295a2a0a2a2a2a8a2aaa219a425a441a444a450a454a4"
b"55a458a45aa461a465a466a468a469a485a406a509a510a512a515a518a526a5"
b"29a542a545a551a554a555a556a559a565a56aa581a584a585a586a589a592a5"
b"95a598a505a611a616a61aa621a625a644a646a64aa652a655a656a658a660a6"
b"62a686a690a695a696a699a6a1a6a4a6a6a600a802a808a80aa820a822a828a8"
b"2aa851a854a856a859a880a882a888a88aa895a8a0a8a2a8a8a8aaa805a914a9"
b"19a921a925a941a950a955a95aa961a966a969a990a996a900aa02aa08aa0aaa"
b"20aa22aa28aa2aaa51aa54aa56aa80aa82aa88aa8aaa95aaa0aaa2aaa8aaaaaa"
)
delta = np.float32(0.125)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d, rest = np.hsplit(blocks, [2])
qs, qh = np.hsplit(rest, [QK_K // 8])
d = d.view(np.float16).astype(np.float32)
qh = qh.view(np.uint16)
dl = d * (2 * ((qh >> 12) & 7) + 1)
dl = dl.reshape((n_blocks, -1, 1, 1))
delta = np.where((qh & np.uint16(0x8000)) == 0, cls.delta, -cls.delta)
delta = delta.reshape((n_blocks, -1, 1, 1))
qh = qh.reshape((n_blocks, -1, 1)) >> np.array([0, 3, 6, 9], dtype=np.uint16).reshape((1, 1, 4))
qs = qs.astype(np.uint16) | ((qh & 7) << 8).reshape((n_blocks, -1))
assert cls.grid is not None
grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
grid = grid.reshape((n_blocks, -1, 4, 8))
return (dl * (grid + delta)).reshape((n_blocks, -1))
class IQ1_M(__Quant, qtype=GGMLQuantizationType.IQ1_M):
grid_shape = IQ1_S.grid_shape
grid_map = IQ1_S.grid_map
grid_hex = IQ1_S.grid_hex
delta = IQ1_S.delta
# Okay *this* type is weird. It's the only one which stores the f16 scales in multiple parts.
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
qs, rest = np.hsplit(blocks, [QK_K // 8])
qh, scales = np.hsplit(rest, [QK_K // 16])
# The f16 scale is packed across multiple bytes
scales = scales.view(np.uint16)
d = (scales.reshape((n_blocks, 4)) & np.uint16(0xF000)) >> np.array([12, 8, 4, 0], dtype=np.uint16).reshape((1, 4))
d = d[..., 0] | d[..., 1] | d[..., 2] | d[..., 3]
d = d.view(np.float16).astype(np.float32).reshape((n_blocks, 1))
scales = scales.reshape(n_blocks, -1, 1) >> np.array([0, 3, 6, 9], dtype=np.uint16).reshape((1, 1, 4))
scales = (scales & 0x07).reshape((n_blocks, -1))
dl = d * (2 * scales + 1)
dl = dl.reshape((n_blocks, -1, 2, 1, 1))
qh = qh.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
qs = qs.astype(np.uint16) | ((qh & 0x07).astype(np.uint16) << 8).reshape((n_blocks, -1))
delta = np.where(qh & 0x08 == 0, cls.delta, -cls.delta)
delta = delta.reshape((n_blocks, -1, 2, 2, 1))
assert cls.grid is not None
grid = np.take_along_axis(cls.grid, qs.reshape((n_blocks, -1, 1, 1)), axis=-2)
grid = grid.reshape((n_blocks, -1, 2, 2, 8))
return (dl * (grid + delta)).reshape((n_blocks, -1))
class IQ4_NL(__Quant, qtype=GGMLQuantizationType.IQ4_NL):
kvalues = (-127, -104, -83, -65, -49, -35, -22, -10, 1, 13, 25, 38, 53, 69, 89, 113)
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d, qs = np.hsplit(blocks, [2])
d = d.view(np.float16).astype(np.float32)
qs = qs.reshape((n_blocks, -1, 1, cls.block_size // 2)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
qs = (qs & np.uint8(0x0F)).reshape((n_blocks, -1, 1))
kvalues = np.array(cls.kvalues, dtype=np.int8).reshape(1, 1, 16)
qs = np.take_along_axis(kvalues, qs, axis=-1).astype(np.float32).reshape((n_blocks, -1))
return (d * qs)
class IQ4_XS(__Quant, qtype=GGMLQuantizationType.IQ4_XS):
@classmethod
def dequantize_blocks(cls, blocks: np.ndarray) -> np.ndarray:
n_blocks = blocks.shape[0]
d, rest = np.hsplit(blocks, [2])
scales_h, rest = np.hsplit(rest, [2])
scales_l, qs = np.hsplit(rest, [QK_K // 64])
d = d.view(np.float16).astype(np.float32)
scales_h = scales_h.view(np.uint16)
scales_l = scales_l.reshape((n_blocks, -1, 1)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2))
scales_h = scales_h.reshape((n_blocks, 1, -1)) >> np.array([2 * i for i in range(QK_K // 32)], dtype=np.uint16).reshape((1, -1, 1))
scales_l = scales_l.reshape((n_blocks, -1)) & np.uint8(0x0F)
scales_h = scales_h.reshape((n_blocks, -1)).astype(np.uint8) & np.uint8(0x03)
scales = (scales_l | (scales_h << np.uint8(4))).astype(np.int8) - np.int8(32)
dl = (d * scales.astype(np.float32)).reshape((n_blocks, -1, 1))
qs = qs.reshape((n_blocks, -1, 1, 16)) >> np.array([0, 4], dtype=np.uint8).reshape((1, 1, 2, 1))
qs = qs.reshape((n_blocks, -1, 32, 1)) & np.uint8(0x0F)
kvalues = np.array(IQ4_NL.kvalues, dtype=np.int8).reshape((1, 1, 1, -1))
qs = np.take_along_axis(kvalues, qs, axis=-1).astype(np.float32).reshape((n_blocks, -1, 32))
return (dl * qs).reshape((n_blocks, -1))