vLLM makes FlashMLA mega-attention the SM100 default for DeepSeek V4.1, with an NVFP4 compressed KV record only that kernel can read

GitHub·high signalXBlueskyLinkedInCopy link

PR #56935 landed FlashMLAMegaAttnBackend (FLASHMLA_MEGA_ATTN_DSV41), which fuses Q RoPE, sparse attention, the output's inverse RoPE and its FP8 cast into one launch writing straight into the buffer wo_a consumes, and as of this revision it is the SM100 default rather than opt-in. It adds an nvfp4_ds_mla KV format: a 288-byte compressed record of 256 B e2m1 pairs plus 32 e4m3 scales over 16 dims, matching what the reference inference/model.py stores. Two orthogonal knobs (apply_q_rope, is_q_interleaved) ride the existing fused insert op, pinned by a bit-for-bit test over all eight combinations of 16/64 heads x norm x RoPE.