Most LLM architectures use “soft attention” where some fractional amount of attention is put on every token. “Hard attention” is the term for what you describe.