I made a kernel 2.2x faster. It made my training loop 3x slower

Keyword: Vllm

Publisher: Bearblog.dev

Published: Jun 2, 2026

02 Jun, 2026 I wrote a fused decode-attention kernel for an RL training loop, got it 2.2× faster than the SDPA path it replaces at the microbenchmark level, dropped it into HuggingFace's generate, a… [+45364 chars]

Read Full Story ↗