---

name: bitsandbytes description: "Use bitsandbytes for k-bit PyTorch quantization, Hugging Face quantized model loading, 8-bit and paged optimizers, direct quantized layers/functions, and backend installation diagnostics." disable-model-invocation: true metadata: disco-role: operating license: MIT

bitsandbytes

Use this repo skill when a task involves bitsandbytes, k-bit PyTorch quantization, LLM.int8(), QLoRA, 4-bit/NF4/FP4 layers, 8-bit optimizers, paged optimizers, or bitsandbytes install/backend failures.

Start Here

1. For any new environment, install with pip install bitsandbytes and verify with:

   python -c "import bitsandbytes as bnb; print(bnb.__version__)"
   python -m bitsandbytes
   

2. If import or backend diagnostics fail, route first to sub-skills/installation-diagnostics/SKILL.md.

3. If the user is using Hugging Face BitsAndBytesConfig, route to sub-skills/transformers-integrations/SKILL.md.

4. If the user is replacing layers or calling bitsandbytes.functional, route to sub-skills/quantized-modules-functions/SKILL.md.

5. If the user is choosing or debugging bitsandbytes.optim, route to sub-skills/optimizers-training/SKILL.md.

Route Map

User asks about	Use	Why
pip install bitsandbytes, import bitsandbytes, python -m bitsandbytes, CUDA/ROCm/XPU/HPU/MPS/CPU support, missing libbitsandbytes_*, source builds, BNB_CUDA_VERSION, BNB_ROCM_VERSION	sub-skills/installation-diagnostics/SKILL.md	Owns install/backend compatibility and native library troubleshooting.
Transformers, Diffusers, PEFT, Accelerate, BitsAndBytesConfig, load_in_8bit, load_in_4bit, NF4, QLoRA, FSDP-QLoRA	sub-skills/transformers-integrations/SKILL.md	Owns Hugging Face model-loading and finetuning integration patterns.
Linear8bitLt, Linear4bit, LinearNF4, Embedding8bit, Params4bit, Int8Params, QuantState, quantize_4bit, int8 vectorwise quantization, direct matmul	sub-skills/quantized-modules-functions/SKILL.md	Owns direct module/function API usage and state-dict caveats.
Adam8bit, AdamW8bit, PagedAdamW8bit, Lion8bit, AdEMAMix8bit, GlobalOptimManager, StableEmbedding, optimizer memory savings	sub-skills/optimizers-training/SKILL.md	Owns optimizer selection, training-loop integration, and state checks.

Shared References and Scripts

• Read references/repo-provenance.md before deciding whether this skill matches a current bitsandbytes checkout or should be refreshed.
• Read references/troubleshooting.md for cross-cutting routing from symptoms to the right sub-skill.
• Read references/installation-compatibility.md for public install requirements and backend support summary.
• Read references/performance-and-benchmarks.md before interpreting memory or speed claims.
• Run scripts/check-bitsandbytes-install.py --json for a safe import/backend report that delegates to the bundled installation diagnostic helper.

Common Decision Points

• BitsAndBytesConfig belongs to Transformers, not to the bitsandbytes package. Use the Transformers integration sub-skill for those configs.
• CPU-only environments can validate imports, signatures, and some construction paths, but they do not prove CUDA/ROCm/XPU kernels or memory savings.
• Direct quantized layers usually quantize when moved to a real device with .to(device). Construction on CPU is not the same as executing quantized kernels.
• 8-bit optimizer memory savings depend on optimizer-state size; small tensors below min_8bit_size=4096 intentionally remain 32-bit.
• Paged optimizers and many quantized model-loading paths need supported accelerator behavior and should not be promised from CPU-only checks.

Safe Validation Commands

python scripts/check-bitsandbytes-install.py --json
python sub-skills/installation-diagnostics/scripts/backend-report.py --json
python sub-skills/quantized-modules-functions/scripts/quantized-module-smoke.py --json
python sub-skills/optimizers-training/scripts/cpu-optimizer-smoke.py --optimizer adam8bit --steps 3
python sub-skills/transformers-integrations/scripts/transformers-bnb-config-check.py --mode qlora --json

Only run GPU or model-loading checks after confirming hardware, optional dependencies, model access, and that downloads or cache use are allowed.