qwen3 tts speed cost frontier
softmax free attention gpt2 medium
tritonsigmoid fast sigmoid attention
nvidia tensorrt dynamo triton
llms are complicated now
kernelgenbench multi source multi chip benchmark