vLLM moves DeepSeek-V4.1 onto FlashMLA's own KV record, cutting 584 bytes per token to 528

GitHub·high signalXBlueskyLinkedInCopy link

Merged 2026-09-15, vLLM PR #56893 switches DeepSeek-V4.1 off the DeepSeek-V4 paged fp8_ds_mla record onto a V4.1-specific one that quantizes the RoPE dims too: 512 B of fp8 e4m3 data plus 16 UE8M0 scales (one per 32 dims, MXFP8), against V4's 448 B fp8 NoPE + 128 B bf16 RoPE and 7 scales. Because block_size is always a multiple of 32, every V4.1 page is an exact multiple of the 512 B TMA stride, so the padding V4 needed disappears on top of the 9.6% smaller record. The subtle part: the DeepSeek-V4 indexer cache's 576 B alignment existed only to keep the summed block stride divisible, and leaving it while the main caches moved produced a stride of 212160 that tripped FlashMLA's assertion after a clean weight load with every per-spec page size correct in isolation.