publication2026Unweight: Lossless MLP Weight Compression for LLM InferenceIvan NikulinMachine LearningCompressionGPU Systems