Any update on this? Our team all have m2-m4 mbp and they do not trust the machines because simple matrix computation give wrong results This is a simple test we did from rich import print import os os.environ["TF_CPP_MIN_LOG_LEVEL"] = "4" import numpy as np import tensorflow as tf RT = tf.constant( [ [-0.25497323, -0.81989247, -0.5126062, 0.3136883], [-0.32365915, 0.57191426, -0.75376326, 0.36354592], [0.9111716, -0.02627973, -0.41118845, 0.511739], [0.0, 0.0, 0.0, 1.0], ] ) def invert(RT): """ I found this bug while applying an inverse transform to a vector and the inverted transform was wrong """ # test data from my dataset R = RT[..., :3, :3] T = RT[..., :3, 3] R_inv = tf.einsum("...ij->...ji", R) T_inv = -tf.einsum("...ij,...j->...i", R_inv, T) # (..., 3) return T_inv print("Numpy Sanity check") print(f"np inv:\n{np.linalg.inv(RT)[:3, 3]}") print(f"np inv Float32:\n{np.linalg.inv(tf.cast(RT, tf.float32))[:3, 3]}") print(f"np inv Float16:\n{np.linalg.inv(tf.cast(RT, tf.float16))}") with tf.device("/GPU:0"): res = invert(RT) print(f"\nTF on {res.device}") print(f"tf inv:\n{res}") print(f"tf inv Float32:\n{invert(tf.cast(RT, tf.float32))}") print(f"tf inv Float16:\n{invert(tf.cast(RT, tf.float16))}") with tf.device("/CPU:0"): res = invert(RT) print(f"\nTF on {res.device}") print(f"tf inv:\n{res}") print(f"tf inv Float32:\n{invert(tf.cast(RT, tf.float32))}") print(f"tf inv Float16:\n{invert(tf.cast(RT, tf.float16))}")

sorry for the mangled example above, I was on my iPhone when posting. I have nailed the culprit, and it is T = RT[..., :3, 3] which on CPU works as expected but not on GPU. using T = tf.reshape(tf.slice(RT, [0, 3], [3, 1]), [3]) produces consistent results on CPU and GPU

I did a few more tests and RT[:3, 3] works correctly while RT[..., :3, 3] does not. Check this minimal repro: import tensorflow as tf RT = tf.constant( [ [-0.25497323, -0.81989247, -0.5126062, 0.3136883], [-0.32365915, 0.57191426, -0.75376326, 0.36354592], [0.9111716, -0.02627973, -0.41118845, 0.511739], [0.0, 0.0, 0.0, 1.0], ] ) with tf.device("/GPU:0"): T1 = RT[..., :3,3] T2 = RT[:3,3] print(f"RT[..., :3,3]: {T1}") print(f"RT[:3,3]: {T2}") output: 2025-10-04 15:16:12.602764: I metal_plugin/src/device/metal_device.cc:1154] Metal device set to: Apple M4 Max 2025-10-04 15:16:12.602789: I metal_plugin/src/device/metal_device.cc:296] systemMemory: 128.00 GB 2025-10-04 15:16:12.602794: I metal_plugin/src/device/metal_device.cc:313] maxCacheSize: 53.76 GB WARNING: All log messages before absl::InitializeLog() is called are written to STDERR I0000 00:00:1759583772.602803 5120482 pluggable_device_factory.cc:305] Could not identify NUMA node of platform GPU ID 0, defaulting to 0. Your kernel may not have been built with NUMA support. I0000 00:00:1759583772.602816 5120482 pluggable_device_factory.cc:271] Created TensorFlow device (/job:localhost/replica:0/task:0/device:GPU:0 with 0 MB memory) -> physical PluggableDevice (device: 0, name: METAL, pci bus id: <undefined>) RT[..., :3,3]: [0.3136883 0.3136883 0.3136883] RT[:3,3]: [0.3136883 0.36354592 0.511739 ] tf 2.19, python 3.12, tf metal 1.2