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

Jump to bottom

2024-05-30 Add FP8 PTQ #1877

Open
wants to merge 8 commits into
base: develop
Choose a base branch
from
Open

2024-05-30 Add FP8 PTQ #1877

Changes from 1 commit
Commits
Show all changes
8 commits
Select commit Hold shift + click to select a range
ce12dc6
2024-05-30 Add FP8 PTQ
YZW-explorer May 30, 2024
64627ba
dequantize_linear quantize_linear now support fp8 ptq function
YZW-explorer Jun 13, 2024
0085711
add fp8 quantization description in post_training_quantization.md
YZW-explorer Jun 14, 2024
6bdcff0
Deleting unnecessary code may lead to runtime conflict errors in FP.
YZW-explorer Jun 18, 2024
4a6c45d
fix overflow bug in base_hist.py
YZW-explorer Jun 18, 2024
c3b56e6
KLObserver now support fp8
YZW-explorer Jun 19, 2024
20bec2b
Add notes in base_hist.py
YZW-explorer Jun 19, 2024
ebca0b9
change docs/zh_cn/tutorials/quant/post_training_quantization.md
YZW-explorer Jun 25, 2024
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
48 changes: 32 additions & 16 deletions paddleslim/quant/observers/uniform.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -56,12 +56,24 @@ def __init__(
def qmin_qmax(self):
""" Calculate the range of the quantized integer based on the specified
quant_bits, sign, and symmetric properties."""
if self._sign:
self._qmin = -2**(self.bit_length() - 1)
self._qmax = 2**(self.bit_length() - 1) - 1
else:
self._qmin = 0
self._qmax = 2**self.bit_length()
if isinstance(self._quant_bits,tuple):
if (self._quant_bits[0]==4 and self._quant_bits[1]==3 and len(self._quant_bits)==2):
self._qmin = -448.0
self._qmax = 448.0
elif (self._quant_bits[0]==5 and self._quant_bits[1]==2 and len(self._quant_bits)==2):
self._qmin = 57344.0
self._qmax = 57344.0
else:
raise NotImplementedError(
"Currently, only float8_e4m3 and float8_e5m2 formats are supported. Please set quant_bits to (4,3) or (5,2) for the corresponding format."
)
else:
if self._sign:
self._qmin = -2**(self.bit_length() - 1)
self._qmax = 2**(self.bit_length() - 1) - 1
else:
self._qmin = 0
self._qmax = 2**self.bit_length()
return self._qmin, self._qmax

@abc.abstractmethod
Expand All @@ -87,15 +99,19 @@ def cal_scales_zero_points(self) -> Tuple[float, float]:
# It is important to ensure that common operations like zero padding do not cause quantization errors.
_min = min(self.min_value(), 0.)
_max = max(self.max_value(), 0.)

if self._symmetric:
self._scale = max(-_min, _max)
if self._sign:
self._zero_point = 0
if isinstance(self._quant_bits,tuple):
_abs_max = max(-_min, _max)
self._scale = _qmax / _abs_max
self._zero_point = 0
else:
if self._symmetric:
self._scale = max(-_min, _max)
if self._sign:
self._zero_point = 0
else:
self._zero_point = (_qmax + _qmin) / 2
else:
self._zero_point = (_qmax + _qmin) / 2
else:
self._scale = (_max - _min) / float(_qmax - _qmin)
self._zero_point = _qmin - round(_min / self._scale)
self._zero_point = np.clip(self._zero_point, _qmin, _qmax)
self._scale = (_max - _min) / float(_qmax - _qmin)
self._zero_point = _qmin - round(_min / self._scale)
self._zero_point = np.clip(self._zero_point, _qmin, _qmax)
return self._scale, self._zero_point