Worked on the fla-org/flash-linear-attention repository to enhance kernel autotuning reliability for Blackwell GPUs, focusing on stability for large-model inference and MoE+Mamba workloads. Addressed a critical bug by registering a default global_scratch allocator, which prevented NullAllocator crashes and deadlocks during autotuning. Refactored capability checks using shared IS_NVIDIA_BLACKWELL logic, improving forward compatibility with future NVIDIA architectures and simplifying ongoing maintenance. Consolidated allocator management to reduce future code churn and support evolving hardware. Leveraged CUDA, Python, and deep learning expertise to ensure stable performance for vLLM with Qwen3-Coder-Next and Qwen3.5 models on next-generation GPU platforms.