Unweight: how we compressed an LLM 22% without sacrificing quality
Unweight is Cloudflare’s lossless compression system for LLM weights, reducing model size by 15–22% while preserving bit-exact outputs. It targets the memory-bandwidth bottleneck in GPU inference by compressing weights in HBM and decompressing them directly into fast on-chip memory before tensor-core computation. On Llama-3.1-8B, the approach saves roughly 3 GB of VRAM and enables more models to run per GPU. ## The GPU Memory Bottleneck - LLM inference is often limited by memory bandwidth rather than computation. - Each generated token requires reading the model’s weights from GPU high-bandwidth memory (HBM). - NVIDIA H100 tensor cores can process data far faster than HBM can supply it. - Smaller weights reduce the amount of data transferred across the memory bus. - Decompression must be carefully integrated: if it adds latency that cannot overlap with matrix multiplication, token generation becomes slower. ## Why Lossless Compression Matters - Quantization commonly converts 16-bit values into 8- or 4-bit integers. - Because quantization is lossy, it can change model behavior and response quality unpredictably. - Unweight instead preserves exact outputs and does not require specialized hardware. - Existing systems were unsuitable because they focused on CPU decompression, custom FPGA hardware, or consumer GPUs rather than Hopper-generation GPUs and production inference. ## Compressing BF16 Weights - BF16 values contain: - A sign bit - An 8-bit exponent - A 7-bit mantissa - Sign and mantissa values appear largely random and are difficult to compress. - Exponents are highly predictable: the 16 most common exponent values account for more than 99% of weights in a typical layer. - Unweight applies Huffman coding to exponent bytes while leaving sign and mantissa bits unchanged. - Rare exponents are handled by storing an entire row of 64 weights verbatim, avoiding per-element branching during decoding. ## Selective Compression of Model Layers - Unweight compresses the MLP gate, up, and down projection matrices. - These matrices represent roughly two-thirds of model parameters and generate substantial memory traffic during decoding. - Attention weights, embeddings, and layer norms remain uncompressed. - The exponent compression produces about 30% savings in the targeted streams and approximately 20% reduction in total MLP weight size. - Overall model-size reductions reach 15–22%. ## Direct GPU Decompression - Model weights normally reside in large but slower HBM and are staged into small, fast shared memory before computation. - Conventional approaches decompress full matrices back into HBM and then run standard matrix multiplication, creating additional memory traffic. - Unweight decompresses weights in shared memory and feeds them directly to tensor cores. - Different execution strategies are used depending on the weight matrix and batch size. - An autotuner selects the fastest strategy for each workload. ## Results and Availability - Tests on Llama-3.1-8B achieved: - Around 30% compression for MLP weights - 15–22% reduction in total model size - Approximately 3 GB of VRAM savings - The savings allow more models to fit on each GPU, potentially reducing inference cost and improving global deployment coverage. - Cloudflare is publishing a technical paper and open-sourcing the GPU kernels. Unweight demonstrates that lossless, inference-time compression can improve GPU utilization without changing model behavior. The practical recommendation is to compress the portions of a model that dominate memory traffic while integrating decoding directly into the GPU execution path.
Read original(opens in new tab)
