The sequence of execution before FastAPI 0.106.0 was like this diagram:

 * If you use masked losses with Keras the loss values may be different in TensorFlow `2.12` compared to previous versions.
 * In cases where the mask was previously ignored, you will now get an error if you pass a mask with an incompatible shape.

### Major Features and Improvements

`tf.keras`:

func gcmCounterCryptGener(b *aes.Block, out, src []byte, counter *[gcmBlockSize]byte) { var mask [gcmBlockSize]byte for len(src) >= gcmBlockSize { aes.EncryptBlockInternal(b, mask[:], counter[:]) gcmInc32(counter) subtle.XORBytes(out, src, mask[:]) out = out[gcmBlockSize:] src = src[gcmBlockSize:] } if len(src) > 0 { aes.EncryptBlockInternal(b, mask[:], counter[:]) gcmInc32(counter) subtle.XORBytes(out, src, mask[:]) } } // gcmInc32 treats the final four bytes of counterBlock as a big-endian value //...

 50% {\n    opacity: 0.2;\n  }\n}\n.placeholder-wave {\n  -webkit-mask-image: linear-gradient(130deg, #000 55%, rgba(0, 0, 0, 0.8) 75%, #000 95%);\n  mask-image: linear-gradient(130deg, #000 55%, rgba(0, 0, 0, 0.8) 75%, #000 95%);\n  -webkit-mask-size: 200% 100%;\n  mask-size: 200% 100%;\n  animation: placeholder-wave 2s linear infinite;\n}\n\n@keyframes placeholder-wave {\n  100% {\n    -webkit-mask-position: -200% 0%;\n    mask-position: -200% 0%;\n  }\n}\n.clearfix::after {\n  display: block;\n...

1F191..1F19A  ; valid                  ;      ; NV8    # 6.0  SQUARED CL..SQUARED VS
1F19B..1F1AC  ; valid                  ;      ; NV8    # 9.0  SQUARED THREE D..SQUARED VOD
1F1AD         ; valid                  ;      ; NV8    # 13.0 MASK WORK SYMBOL
1F1AE..1F1E5  ; disallowed                             # NA   <reserved-1F1AE>..<reserved-1F1E5>