PyTorch 2.15 regression slows LU solver up to 100x, warns researcher
ducha_aiki · x · 2026-11-12
Researcher duchaaiki warns against upgrading to PyTorch 2.15: torch.linalg.lufactorex regresses up to 100x. Benchmarks on RTX 4090 with 8192 (9,8) float32 matrices show 8ms on 2.14.0+cu130 vs 924ms on 2.15.0+cu132 — 2.14 picked a batched MAGMA kernel, but 2.15 forces looped speed on every backend, MAGMA included. Downstream, Kornia's RANSAC fundamental-matrix calls go from 42ms to 1.08s and homography calls from 125ms to 0.98s, in some settings 3.5-80x slower than CPU. The slowdown stems from code changes between 2.14 and 2.15, not CUDA 13 builds.
More from coding & agent
- Broke student's vibe coding stack: ChatGPT writes the plan.md, free agents do the work — lo_siento0 · 2026-11-12
- Local Agav CLI asks permission before any file write, praised for blocking destructive commands — raw_ofc1 · 2026-11-12
- Make every agent action idempotent before letting it retry anything — useralltime_835 · 2026-11-12
- antirez: AI code generation now works when task boundaries are clear — antirez · 2026-11-12
- One-prompt 3D Solar System: Qoder's new Qwen4-Max-Preview builds an explorable browser universe — hey_abusiddik · 2026-11-12
- Giving agents a whiteboard: Claude Code crew plans on a shared Excalidraw canvas for $0.36 — pramodk73 · 2026-11-12