Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Fix scale calculation #2788

Closed
wants to merge 1 commit into from
Closed

Fix scale calculation #2788

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 10 additions & 14 deletions nncf/quantization/algorithms/weight_compression/weight_lowering.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -253,23 +253,19 @@ def calculate_integer_quantization_params(
if weight.dtype != TensorDataType.float32:
weight = weight.astype(TensorDataType.float32)

if mode in [CompressWeightsMode.INT8_ASYM, CompressWeightsMode.INT4_ASYM]:
level_low = 0
level_high = 2**num_bits - 1
asym_quant = mode in [CompressWeightsMode.INT8_ASYM, CompressWeightsMode.INT4_ASYM]
level_low = 0 if asym_quant else -(2 ** (num_bits - 1))
level_high = 2**num_bits - 1 if asym_quant else 2 ** (num_bits - 1) - 1

if asym_quant:
min_values = fns.min(weight, axis=reduction_axes, keepdims=True) # [a1, r, a2] -> [a1, 1, a2]
max_values = fns.max(weight, axis=reduction_axes, keepdims=True) # [a1, r, a2] -> [a1, 1, a2]
scale, zero_point = calculate_scale_zero_point(
min_values, max_values, level_low, level_high, narrow_range=False
)
return scale, zero_point
else:
max_values = fns.max(fns.abs(weight), axis=reduction_axes, keepdims=True) # [a1, r, a2] -> [a1, 1, a2]
min_values = -max_values

level_high = 2 ** (num_bits - 1) - 1
scale = fns.max(fns.abs(weight), axis=reduction_axes, keepdims=True) # [a1, r//gs, 1, a2]
scale /= level_high
eps = fns.finfo(scale).eps
# NOTE: adding machine epsilon to avoid division by zero
scale = fns.where(fns.abs(scale) < eps, eps, scale)
return scale, None
scale, zero_point = calculate_scale_zero_point(min_values, max_values, level_low, level_high, narrow_range=False)
return scale, zero_point if asym_quant else None


def calculate_quantized_weight(
Expand Down
16 changes: 16 additions & 0 deletions tests/openvino/native/quantization/test_weights_compression.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -28,6 +28,7 @@
from nncf.quantization.algorithms.weight_compression.config import WeightCompressionConfig
from nncf.quantization.algorithms.weight_compression.config import WeightCompressionParameters
from nncf.quantization.algorithms.weight_compression.mixed_precision import MIXED_PRECISION_CRITERIA
from nncf.quantization.algorithms.weight_compression.weight_lowering import do_integer_quantization
from nncf.quantization.algorithms.weight_compression.weight_lowering import get_integer_quantization_error
from nncf.quantization.algorithms.weight_compression.weight_lowering import reshape_weight_for_grouped_quantization
from nncf.scopes import IgnoredScope
Expand Down Expand Up @@ -912,3 +913,18 @@ def test_mixed_precision_e2m1(mode, all_layers, ratio, ref_ids):
}
ref_e8m0_nodes = {f"weights_{i}/scale" for i in ref_ids}
assert ref_e8m0_nodes == names_e8m0


def test_compressed_weighs_range():
bits = 4
sz = 2 ** (bits - 1)
quantized_w = np.arange(-sz, sz).reshape(2, sz).astype(np.float32)
w = Tensor(quantized_w / 10.0)

config = WeightCompressionConfig(mode=CompressWeightsMode.INT4_SYM)
compressed_weighs, scale, zp = do_integer_quantization(w, -1, config)

assert zp is None
ref_scale = 2 * np.max(np.abs(quantized_w) / 10.0, axis=1) / (2**bits - 1)
assert np.allclose(scale.data, ref_scale.reshape(scale.shape))
assert np.allclose(compressed_weighs.data, quantized_w)
Loading