Practical experience in writing or optimizing custom kernels ( Pallas , Triton , CUDA , CUTLASS ), performance profiling and trace analysis ( JAX/XLA profiler , Kineto , torch.profiler , Nsight ), debugging distributed or accelerator-bound workloads, serving large language models at scale ( vLLM , SGLang , TensorRT-LLM , JetStream , KV cache , paged attention, continuous batching). Framework-level depth is a strong plus: custom operators, sharding and distributed training ( pjit/shard_map , FSDP , Megatron , MaxText ), or compiler and graph-level work ( XLA/StableHLO , torch.compile , MLIR ).