During March 2025, Aniz contributed to the Furion-cn/sglang repository by implementing BF16 dequantization support for quantized weights, refactoring the dequantization kernel to handle both float16 and bf16 outputs. Using C++ and CUDA, Aniz enhanced the kernel’s precision handling and updated the test suite to validate performance across diverse hardware configurations. Additionally, Aniz addressed a compatibility issue between DeepGemm and Torch Compile by introducing a capability check that conditionally registers the DeepGemm FP8 kernel, ensuring correct kernel selection per system. This work improved numerical accuracy, reduced runtime errors, and strengthened the robustness of quantization workflows in deep learning applications.