January 2026 monthly summary for huggingface/diffusers: Implemented device-aware optimization in the LongCat Image Pipeline by offloading and quantizing the text_encoder to the appropriate device. This reduced cross-device data transfers, improved inputs processing efficiency, and boosted image generation throughput. The change ensures generated_ids are moved instead of the text_encoder, simplifying device placement and reducing bottlenecks. Key commit b351be2379ed14f8599301a69739d7e59e220de1 documents the approach and rationale. Overall impact: faster, more scalable LongCat inference on heterogeneous hardware, enabling more responsive applications and better hardware utilization. Technologies demonstrated include PyTorch device placement, quantization, and inference optimizations within the diffusers codebase.