transformers

c73f3d33 - FIX CIS: whisper encoder casts input_features to conv1 dtype

FIX CIS: whisper encoder casts input_features to conv1 dtype The feature extractor emits float32 mel-spectrograms regardless of model dtype. When the model is loaded in fp16/bf16 the conv1 call raised 'Input type (float) and bias type (c10::Half) should be the same' — the trace for every Whisper fp16 integration test on CI. Cast input_features to the conv1 weight dtype before the first conv to align the two. Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>