LLM 推理链路优化 — Prefill/Decode 基准、KV Cache、GQA/MLA、Spec Decoding（v0.2.0）
numpy pytorch transformer llama attention-mechanism rotary self-attention prefill inference-optimization gqa kv-cache llm rmsnorm llm-inference grouped-query-attention speculative-decoding deepseek swiglu multi-head-latent-attention decode-acceleration
-
Updated
Jul 28, 2026 - Python