Skip to content
New issue

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

#

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

Already on GitHub? # to your account

Fix (ptq): fix for residual with mha #681

Merged
merged 1 commit into from
Jul 18, 2023
Merged

Fix (ptq): fix for residual with mha #681

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
13 changes: 10 additions & 3 deletions src/brevitas/graph/quantize_impl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -16,6 +16,7 @@
ADD_FNS = [torch.add, operator.add, operator.iadd]

ADD_METHODS = ['add', 'add_']

CAT = brevitas.original_cat

SIGN_PRESERVING_MODULES = (
Expand Down Expand Up @@ -46,6 +47,8 @@
nn.PixelUnshuffle,
nn.Identity)

MAX_RESIDUAL_ITERS = 9999


def inp_placeholder_handler(model, input_quantizer):
"""
Expand Down Expand Up @@ -253,8 +256,7 @@ def recursive_input_handler(
else:
assert align_output is None, f"align_output {str(align_output)} not supported."
elif inp_node.op == 'call_function' and inp_node.target in [
torch.flatten, torch.reshape, torch.transpose, operator.getitem,
operator.__getitem__]:
torch.flatten, torch.reshape, torch.transpose]:
recursive_input_handler(
model,
inp_node,
Expand Down Expand Up @@ -307,6 +309,7 @@ def _get_quant_module(model, node, quant_identity_map, quant_act_map, unsigned_a

def residual_handler(
model, quant_identity_map, quant_act_map, unsigned_act_tuple, align_input_quant_fn):
iter = 0

def is_converged(model):

Expand Down Expand Up @@ -349,7 +352,11 @@ def is_converged(model):
return True

while not is_converged(model):
continue
iter += 1
if iter == MAX_RESIDUAL_ITERS:
raise RuntimeError(
"Residual handler could not find a solution to align scale factors "
"across ADDs and CATs")

return model

Expand Down