llama.cpp/gguf-py/tests/test_quants.py
compilade 9bc6db28d0
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.
2024-09-05 21:48:47 -04:00

239 lines
9.8 KiB
Python
Executable File

#!/usr/bin/env python3
# Test gguf.quants so that it exactly matches the C implementation of the (de)quantization
# NOTE: this is kind of a mess, but at least it worked for initially testing the Python implementations.
from __future__ import annotations
import argparse
from math import prod
import os
import sys
from pathlib import Path
import ctypes
import logging
import numpy as np
# Necessary to load the local gguf package
if "NO_LOCAL_GGUF" not in os.environ and (Path(__file__).parent.parent.parent / 'gguf-py').exists():
sys.path.insert(0, str(Path(__file__).parent.parent))
import gguf
from gguf.constants import GGMLQuantizationType
logger = logging.getLogger("test-quants")
c_float_p = ctypes.POINTER(ctypes.c_float)
class ggml_init_params(ctypes.Structure):
_fields_ = [
("mem_size", ctypes.c_size_t),
("mem_buffer", ctypes.c_void_p),
("no_alloc", ctypes.c_bool),
]
class GGMLQuants:
libggml: ctypes.CDLL
def __init__(self, libggml: Path):
self.libggml = ctypes.CDLL(str(libggml))
self.libggml.ggml_quantize_chunk.restype = ctypes.c_size_t
# enum ggml_type type,
# const float * src,
# void * dst,
# int64_t start,
# int64_t nrows,
# int64_t n_per_row,
# const float * imatrix) {
self.libggml.ggml_quantize_chunk.argtypes = (
ctypes.c_int,
ctypes.POINTER(ctypes.c_float),
ctypes.c_void_p,
ctypes.c_int64,
ctypes.c_int64,
ctypes.c_int64,
ctypes.POINTER(ctypes.c_float),
)
self.libggml.ggml_quantize_requires_imatrix.restype = ctypes.c_bool
self.libggml.ggml_quantize_requires_imatrix.argtypes = (ctypes.c_int,)
for t in (
"q4_0", "q4_1", "q5_0", "q5_1", "q8_0",
"q2_K", "q3_K", "q4_K", "q5_K", "q6_K",
"tq1_0", "tq2_0",
"iq2_xxs", "iq2_xs", "iq2_s", "iq3_xxs", "iq3_s", "iq1_s", "iq1_m",
"iq4_nl", "iq4_xs",
):
dequant_func: ctypes._NamedFuncPointer = getattr(self.libggml, "dequantize_row_" + t)
dequant_func.restype = None
dequant_func.argtypes = (ctypes.c_void_p, ctypes.POINTER(ctypes.c_float), ctypes.c_int64)
self.libggml.ggml_fp16_to_fp32_row.restype = None
self.libggml.ggml_fp16_to_fp32_row.argtypes = (ctypes.POINTER(ctypes.c_uint16), ctypes.POINTER(ctypes.c_float), ctypes.c_int64)
self.libggml.ggml_bf16_to_fp32_row.restype = None
self.libggml.ggml_bf16_to_fp32_row.argtypes = (ctypes.POINTER(ctypes.c_uint16), ctypes.POINTER(ctypes.c_float), ctypes.c_int64)
self.libggml.ggml_init.argtypes = (ggml_init_params,)
self.libggml.ggml_init(ggml_init_params(1 * 1024 * 1024, 0, False))
def dequantize(self, tensor: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray:
result = np.zeros(gguf.quant_shape_from_byte_shape(tensor.shape, qtype), dtype=np.float32, order="C")
if qtype == GGMLQuantizationType.F32:
# no-op
result = tensor.view(np.float32)
elif qtype == GGMLQuantizationType.F16:
self.libggml.ggml_fp16_to_fp32_row(tensor.ctypes.data_as(ctypes.POINTER(ctypes.c_uint16)), result.ctypes.data_as(c_float_p), result.size)
elif qtype == GGMLQuantizationType.BF16:
self.libggml.ggml_bf16_to_fp32_row(tensor.ctypes.data_as(ctypes.POINTER(ctypes.c_uint16)), result.ctypes.data_as(c_float_p), result.size)
else:
lw_qname = qtype.name.lower()
if lw_qname[-1] == "k":
lw_qname = lw_qname[:-1] + "K"
dequant_func: ctypes._NamedFuncPointer = getattr(self.libggml, "dequantize_row_" + lw_qname)
dequant_func(tensor.ctypes.data_as(ctypes.c_void_p), result.ctypes.data_as(c_float_p), result.size)
return result
def quantize(self, data: np.ndarray, qtype: GGMLQuantizationType) -> np.ndarray:
result = np.zeros(gguf.quant_shape_to_byte_shape(data.shape, qtype), dtype=np.uint8, order="C")
if self.libggml.ggml_quantize_requires_imatrix(qtype.value):
# TODO: is a column-wise sum of squares appropriate?
qw = np.sum((data * data).reshape((-1, data.shape[-1])), axis=0).ctypes.data_as(c_float_p)
else:
qw = ctypes.cast(0, c_float_p)
result_size = self.libggml.ggml_quantize_chunk(qtype.value, data.ctypes.data_as(c_float_p), result.ctypes.data_as(ctypes.c_void_p), 0, prod(data.shape[:-1]), data.shape[-1], qw)
assert result.size == result_size
return result
def compare_tensors(t1: np.ndarray, t2: np.ndarray, qtype: GGMLQuantizationType) -> bool:
same = np.array_equal(t1, t2)
if same:
return True
else:
block_size, type_size = gguf.GGML_QUANT_SIZES[qtype]
if t1.dtype == np.float32:
t1 = t1.reshape((-1, block_size))
t2 = t2.reshape((-1, block_size))
else:
t1 = t1.reshape((-1, type_size))
t2 = t2.reshape((-1, type_size))
x = t1.view(np.uint8) ^ t2.view(np.uint8)
diff_bits = np.count_nonzero(np.unpackbits(x, axis=-1), axis=-1)
num_bad_blocks = np.count_nonzero(diff_bits, axis=0)
if num_bad_blocks == 0 and t1.shape == t2.shape:
logger.debug("Bits are equal, but arrays don't match, likely contains NANs")
return True
logger.debug(f"{num_bad_blocks} bad blocks ({100 * num_bad_blocks / x.shape[0]:.6f}%)")
bad_block_id = np.argmax(diff_bits, axis=0)
logger.debug(f"Worst block id: {bad_block_id}")
logger.debug(f"Sample bad block ({diff_bits[bad_block_id]} differing bits):\n{t1[bad_block_id]}\nReference:\n{t2[bad_block_id]}")
sum_diff_bits = np.sum(diff_bits)
logger.debug(f"{sum_diff_bits} bits differ ({100 * sum_diff_bits/(x.size * 8):.6f}%)")
return False
def do_test(libggml_path: Path, quick: bool = False):
ggml_quants = GGMLQuants(libggml_path)
np.set_printoptions(precision=None, threshold=(4 * 256) + 1, formatter={"int": lambda n: "0x%02X" % n})
r = np.random.randn(8, 1024, 1024).astype(np.float32, copy=False)
for qtype in (GGMLQuantizationType.F16, *gguf.quants._type_traits.keys()):
has_dequantize = False
has_quantize = False
try:
gguf.dequantize(np.zeros((gguf.GGML_QUANT_SIZES[qtype][1]), dtype=np.uint8), qtype)
has_dequantize = True
except (NotImplementedError, AssertionError) as e:
if isinstance(e, AssertionError):
logger.error(f"Error with {qtype.name}: {e}")
raise e
try:
gguf.quantize(np.zeros((gguf.GGML_QUANT_SIZES[qtype][0]), dtype=np.float32), qtype)
has_quantize = True
except (NotImplementedError, AssertionError) as e:
if isinstance(e, AssertionError):
logger.error(f"Error with {qtype.name}: {e}")
raise e
if not has_dequantize and not has_quantize:
continue
logger.info(f"Testing {qtype.name}")
rc = r.copy(order="C")
pyq = None
ggq = None
if has_quantize:
logger.debug(f"Quantizing to {qtype.name} with Python")
pyq = gguf.quants.quantize(rc, qtype)
logger.debug(f"Quantizing to {qtype.name} with C")
ggq = ggml_quants.quantize(rc, qtype)
if qtype == GGMLQuantizationType.F16:
pyq = pyq.view(np.uint8)
quant_equal = compare_tensors(pyq, ggq, qtype)
if not quant_equal:
logger.error(f"Quantization to {qtype.name} does not match ❌")
else:
logger.info(f"Quantization to {qtype.name} matches exactly ✅")
if has_dequantize:
if ggq is None and not quick:
logger.debug(f"Quantizing to {qtype.name} with C")
ggq = ggml_quants.quantize(rc, qtype)
if ggq is not None:
logger.debug(f"Dequantizing from {qtype.name} with Python")
pydq = gguf.quants.dequantize(ggq, qtype)
logger.debug(f"Dequantizing from {qtype.name} with C")
ggdq = ggml_quants.dequantize(ggq, qtype)
dequant_equal = compare_tensors(pydq, ggdq, qtype)
if not dequant_equal:
logger.error(f"Dequantization from {qtype.name} does not match ❌")
else:
logger.info(f"Dequantization from {qtype.name} matches exactly ✅")
rq_shape = gguf.quants.quant_shape_to_byte_shape((8, 1024, 1024 // 2), qtype)
rq = np.random.random(rq_shape).astype(np.float16).view(np.uint8)
logger.debug(f"Dequantizing random f16 data as {qtype.name} with Python")
pydq = gguf.quants.dequantize(rq, qtype)
logger.debug(f"Dequantizing random f16 data as {qtype.name} with C")
ggdq = ggml_quants.dequantize(rq, qtype)
dequant_equal = compare_tensors(pydq, ggdq, qtype)
if not dequant_equal:
logger.error(f"Dequantization from random f16 data as {qtype.name} does not match ❌")
else:
logger.info(f"Dequantization from random f16 data as {qtype.name} matches exactly ✅")
if __name__ == "__main__":
parser = argparse.ArgumentParser(description="Test Python (de)quantization against the reference C implementation")
parser.add_argument("--libggml", type=Path, default=Path(__file__).parent.parent.parent / "build" / "ggml" / "src" / "libggml.so", help="The path to libggml.so")
parser.add_argument("--quick", action="store_true", help="Don't quantize with C when it's not strictly necessary")
args = parser.parse_args()
logging.basicConfig(level=logging.DEBUG)
do_test(args.libggml, args.quick)