FlashMLA ported to consumer Blackwell sm_120, 2-3x faster MLA attention vs SDPA
smashedshanky · reddit · 2026-08-30
While building an open-source LLM training repo supporting arbitrary architectures (GQA/MLA/Dense), a developer found DeepSeek's FlashMLA kernel only compiled for data-center sm100/sm90, so they ported and open-sourced a Windows/Linux build for consumer Blackwell sm120.
Inference vs PyTorch SDPA: sparse fp8 decode 2.62x faster; sparse serving 5x; sparse prefill (sq=512, skv=8192) 2.61x; FP8 KV cache gives -8% latency and 1.84x less cache memory (SDPA: +1.8%). Model-level bf16 decode roughly at parity.
Training (fwd+bwd): Dense seq 1024/4096/8192 at 3.29x / 2.40x / 3.04x, sparse prefill 3.01x.
More from Infra
- NVIDIA launches RTX Spark for local AI on Windows PCs — ostrisai · 2026-10-16
- SkyRL ships Day-0 RL support for Beam with +62% LoRA decode throughput in vLLM — vllm_project · 2026-10-16
- DFlash2 delivers up to 3.03x decode-only speedup for Beam across five benchmarks — vllm_project · 2026-10-16
- vLLM's Beam serving kernels cut GB200 forward time from 8.15 to 6.47 ms at TP8 batch 128 — vllm_project · 2026-10-16
- Reflection's 501B MoE Beam hits 463 tok/s on 8x GB200 with NVFP4 at day 0 — zhyncs42 · 2026-10-16
- Gemma 4 31B QAT on one 24GB card: 46 tok/s plain, ~80 with draft model on RTX 4090 — Subject_Poet_3861 · 2026-10-16