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

[Cherry-Pick]Cp fit paddle26 #1823

Merged
merged 5 commits into from
Dec 28, 2023
Merged

[Cherry-Pick]Cp fit paddle26 #1823

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
6bfae6f
upadte
RachelXu7 Dec 28, 2023
ffbf9ba
upadte
RachelXu7 Dec 28, 2023
587d3a8
upadte
RachelXu7 Dec 28, 2023
4dea1a8
Fit for Paddle2.6
RachelXu7 Dec 28, 2023
23690df
cherry-pick
RachelXu7 Dec 28, 2023
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
19 changes: 13 additions & 6 deletions paddleslim/quant/advanced/gptq.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -106,8 +106,9 @@ def fasterquant(self,
H = self.hessian
del self.hessian
dead = paddle.where(paddle.diag(H) == 0)
H[dead, dead] = 1
W[:, dead] = 0
if dead[0].shape[0] != 0:
H[dead, dead] = 1
W[:, dead] = 0
del dead
if actorder:
perm = paddle.argsort(paddle.diag(H), descending=True)
Expand All @@ -122,9 +123,15 @@ def fasterquant(self,
damp = percdamp * paddle.mean(paddle.diag(H))
diag = paddle.arange(self.columns)
H[diag, diag] += damp

H = paddle.inverse(H)
H = paddle.linalg.cholesky(H, upper=True)
try:
H = paddle.inverse(H)
H = paddle.linalg.cholesky(H, upper=True)
except:
print('We skip GPTQ this layer now.')
print(
'If you want GPTQ this layer, please try setting damp_percent larger or increasing the number of samples.'
)
return
Hinv = H

for i1 in range(0, self.columns, blocksize):
Expand Down Expand Up @@ -182,4 +189,4 @@ def fasterquant(self,

self.quantized = True
del H, Q, Hinv, W, Losses
paddle.device.cuda.empty_cache()
paddle.device.cuda.empty_cache()
3 changes: 3 additions & 0 deletions paddleslim/quant/advanced/piecewise_search.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -97,6 +97,8 @@ def search(self, layer_name, sampled_input, act_abs_max, weight):
mask_for_search = paddle.where(labels == centroids.argsort()[i],
1., 0.)
mask_for_ones = paddle.where(mask_for_search == 0., 1., 0.)
mask_for_search = mask_for_search.cast(dtype)
mask_for_ones = mask_for_ones.cast(dtype)

while alpha <= alpha_max:
if alpha < 1:
Expand Down Expand Up @@ -125,6 +127,7 @@ def search(self, layer_name, sampled_input, act_abs_max, weight):
if smooth_scale_out is not None:
mask_for_ones_new = paddle.where(
smooth_scale_out == 0., 1., 0.)
mask_for_ones_new = mask_for_ones_new.cast(dtype)
mask_for_ones *= mask_for_ones_new
smooth_scale_ = smooth_scale_out + smooth_scale
smooth_scale_tmp = smooth_scale_ + mask_for_ones
Expand Down