FP8 GEMM Autotuning: Corrected L2 cache contamination and latency measurement errors in SGLang's block-wise GEMM tuner, improving how it selects kernel configurations. The resulting configuration achieved 7.7% lower kernel latency on a controlled RTX 4090 workload.
GPU Memory Safety: Diagnosed FlashInfer's Gated DeltaNet prefill memory corruption under CUDA MPS and developed a fix that prevents out-of-bounds tensor-map writes under reduced GPU allocations. Built a cached device-level probe to size workspaces from hardware SM-ID bounds while preserving CUDA graph capture compatibility.
Persistent Kernel Caching: Implemented CuTe-DSL disk caching for FlashInfer's MXFP8 GEMM runners, allowing compiled kernels to survive process restarts. Reduced first-call time from 330 ms to 40 ms on a Blackwell warp split-K workload. Added architecture-specific cache keys, source-based invalidation, and automatic recompilation when cached kernels cannot be reused.
MoE Tuner Reliability: Fixed nullable quantization metadata parsing in SGLang's Triton MoE tuner, preventing configuration errors and correctly distinguishing channel, tensor, block, and grouped quantization recipes.