Date: 28.08.2025

BitsAndBytes CUDA Compatibility — quick cheat sheet

Quick reference for BNB - CUDA Toolkit - Compute Capability - PyTorch.
PyTorch versions are approximate and reflect practical compatibility. It’s often simpler to build from source especially for old GPU.

BNB Version	CUDA Toolkit	Min CUDA CC	PyTorch Version
0.43.3	≥11.1	Kepler CC = 3.7	2.2.0 +
0.44.1	≥11.1	Kepler ≈ CC ≥3.5; LLM.int8() 7.5+	2.2.0 +
0.45.5	≥11.7	Maxwell 5.0+ (Kepler dropped since 0.45.0); LLM.int8() 7.5+	2.4 – 2.5
0.46.1	12 +	Turing 7.5+ (Maxwell dropped since 0.46.0)	2.5 +

Notes

• Kepler (CC ~3.5–3.7): prebuilt wheels are long gone; use source builds and Kepler presets. Possible on 0.43.x–0.44.x; starting with 0.45.0 Kepler is removed.
• Maxwell (CC ≥5.0): supported for 8‑bit optimizers and 4‑bit NF4/FP4; marked deprecated in newer branches. LLM.int8() still needs CC ≥7.5 (Turing+).
• PyTorch 2.6: bitsandbytes ≥ 0.45.2 includes a compatibility fix for Triton 3.2.
• Mismatched CUDA between PyTorch and BNB: you can override BNB’s toolkit with BNB_CUDA_VERSION=12x and extend LD_LIBRARY_PATH accordingly.

Quick environment check

python3 -m bitsandbytes