Databricks Proteus Boosts Qwen 3.5 Kernels by 5.2x

Databricks has introduced Proteus, an automated system designed to generate highly specialized GPU kernels. When applied to the Gated DeltaNet path of the Qwen 3.5 122B model running on NVIDIA B200 GPUs with a Triton backend, Proteus produced individual kernels that performed 1.8x to 5.2x faster than the top implementations available in vLLM. Rather than relying on generic kernels that handle diverse workloads suboptimally, Proteus specializes kernels to the exact shapes encountered during runtime, such as fluctuating token counts.

To prevent AI agents from reward-hacking—where they optimize for benchmark metrics rather than genuine performance—Proteus utilizes a strict validation harness. The system tests candidate kernels against a controlled reference implementation, clears leftover compiled states, and uses multiple timers like CUDA event, wall clock, and CUPTI timers. In a case study of a packed decode kernel, which had a baseline latency of 0.025 ms, Proteus rejected incorrect or slower candidates like Candidate 0000. It then split the search into shape-specific paths, yielding Candidate 012 with a 1.5x speedup for single-batch decode, Candidate 030 with a 0.018 ms latency, and Candidate 036 with a 1.6x speedup specialized for a batch size of four and key-value layouts of 128 processed in 64-wide chunks.

A major innovation in Proteus is its streamlined knowledge layer, which manages the trade-off of prompt context. Instead of bloating prompts with costly, distracting historical data, Proteus uses hybrid search and hierarchical tag filtering to deliver only high-trust, actionable takeaways. For machine learning practitioners, this development shifts the bottleneck of kernel optimization from manual programming to automated validation. By allowing autonomous agents to write code while a strict outer loop handles verification and memory, developers can deploy highly efficient, model-specific kernels without risking compiler bugs or false performance metrics.