FlyDSL Kernels Land in PyTorch, Up to 3.01x vLLM Speedups on AMD MI350
PyTorch · x · 2026-10-22
AMD and Meta announced FlyDSL, an open-source Python kernel DSL, now integrates with PyTorch on AMD MI350-series GPUs via torch.compile and eager mode. Kernel-level gains include 1.68x geomean speedups on MXFP4 scaled GEMM and 4.82x on small-K TopK; end-to-end vLLM A/B tests show whole-request speedups of 1.13x (BF16), 2.04x (MXFP8), and 3.01x (MXFP4). Attention kernels ship as a new AOTriton backend.
More from Infra
- CXMT's market cap tops all mid-size Asian fabs combined, half of SK Hynix — pstAsiatech · 2026-10-22
- SemiAnalysis: DeepSeek's open-sourced DeepInfer lacks key inference-serving components — cHHillee · 2026-10-22
- Self-hosted AI rack costs ~$100/mo in power, plus a Bay Area 30¢/kWh problem — mertdumenci · 2026-10-22
- Recursive Ships Flint: a 39.8B MoE Coding Model With 2.5B Active Params and Its Own Inference Engine — BerryBlush_Rogers691 · 2026-10-22
- Meta set to post 23.6% revenue growth to $63.3B in Q3, but capex seen jumping 115% — Beth_Kindig · 2026-10-22
- Free RTX 3090 Speedup: gpu-unlock Boosts Prompt Processing 22-30% in User Tests — Striking_Luck_886 · 2026-10-22