April 2026 monthly summary for pytorch/executorch: Stabilized the quantization path for DecomposeConcatenate to improve FP16 compatibility and robustness in multi-input scenarios. Implemented a targeted bug fix that separates kwargs for quantize_per_tensor and dequantize_per_tensor, ensuring out_dtype is not passed to quantize_per_tensor while preserved for dequantize_per_tensor. This reduces failures in fp16-quantized models and enhances reliability when concatenating inputs that require quantization/dequantization pairs. The change is small but removes a class of edge-case failures that previously affected production deployments.

March 2026 monthly summary for huggingface/transformers: Focused on strengthening model serialization reliability for quantized weights and dequantization paths. Delivered a targeted fix in the dequantization save pathway that enables save_pretrained to persist dequantized weights without conversion, addressing a critical failure point and improving downstream deployment confidence.