Skip to content

Fix duplicated weights in fp8 quantization #37667

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

Merged
Cyrilvallez merged 4 commits into from
Apr 22, 2025
Merged

Fix duplicated weights in fp8 quantization #37667

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
0c372bf
fix fp8
Cyrilvallez Apr 22, 2025
c1c949a
Update quantizer_finegrained_fp8.py
Cyrilvallez Apr 22, 2025
f27e785
fix circular import
Cyrilvallez Apr 22, 2025
4e865a3
Update quantizer_finegrained_fp8.py
Cyrilvallez Apr 22, 2025
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
11 changes: 5 additions & 6 deletions src/transformers/quantizers/quantizer_finegrained_fp8.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -91,11 +91,9 @@ def create_quantized_param(
"""
Quantizes weights to FP8 format using Block-wise quantization
"""
from accelerate.utils import set_module_tensor_to_device
from ..modeling_utils import _load_parameter_into_model

set_module_tensor_to_device(model, param_name, target_device, param_value)

module, tensor_name = get_module_from_name(model, param_name)
param_value = param_value.to(target_device)

# Get FP8 min/max values
fp8_min = torch.finfo(torch.float8_e4m3fn).min
Expand Down Expand Up @@ -131,8 +129,9 @@ def create_quantized_param(
# Reshape scale to match the number of blocks
scale = scale.reshape(scale_orig_shape).squeeze().reciprocal()

module._buffers[tensor_name] = quantized_param.to(target_device)
module._buffers["weight_scale_inv"] = scale.to(target_device)
# Load into the model
_load_parameter_into_model(model, param_name, quantized_param)
_load_parameter_into_model(model, param_name.rsplit(".", 1)[0] + ".weight_scale_inv", scale)

def check_quantized_param(
self,
Expand Down
Loading