[Perf] SM8x sparse-MLA prefill fallback re-reads the shared MLA latent once per head (~64x redundant KV traffic)
![[Perf] SM8x sparse-MLA prefill fallback re-reads the shared MLA latent once per head (~64x redundant KV traffic)](https://www.chat-gpts.plus/wp-content/uploads/2026/09/57971-715139ff-768x403.jpg)
当你看到 [Perf] SM8x sparse-MLA prefill fallback re-reads the shared MLA latent once per head (~64x redundant KV traffic) 这类性能描述时,通常发生在 Ampere(sm_80)显卡上用








