FK Results Differ from Blender

[22TonyFStark]

Hi, thanks for your great code, it helps a lot!

I met a strange question, you can reproduce this situation using the attached files.

I have a target pose A (a simple T-pose, in quaternion form, with a shape of [66, 4], representing 66 bones). I obtained a pose A (in quaternion form) through an optimization method.

When I used pymotion.ops.forward_kinematics.fk to check, I found that the joints obtained from these two poses were very close, and they also looked very similar visually.

To reproduce this, simply run "python debug_train.py"
This code should produce: 2.461019589902293e-07, which is the joint position error, seems good.

However, when I assigned these two poses to the corresponding bones in Blender, the results were significantly different.
To verify this, try the blender file in the attached file.

Attached file:
issue_files.zip

The quaternion above are all normalized already.
What could be the possible reasons for this? Thanks!