Model Quantization Quiz

Answers and explanations appear as you take the quiz.

1. Quantizing a model mainly reduces what?
   • The numerical precision used to store weights and activations
   • The number of layers in the network
   • The training dataset the model was built on
   • The language the model can understand
2. Ignoring scale metadata and other runtime tensors, what is the weight-payload ratio from sixteen bits to four bits per weight?
   • The payload is unchanged
   • It is one hundredth of the original payload
   • The four-bit payload is one quarter of the sixteen-bit payload
   • The payload becomes four times larger
3. What is the main downside of aggressive low-bit quantization?
   • The model forgets its training language
   • Potential loss of accuracy as values are squeezed into fewer buckets
   • It can only run on supercomputers
   • It permanently deletes layers
4. How does quantization-aware training differ from post-training quantization?
   • It requires no data at all
   • It only works on images
   • It increases the number of bits over time
   • It simulates rounding during training so the model learns to tolerate low precision
5. Which technique uses 4-bit quantization to make fine-tuning large models affordable on one GPU?
   • Dropout
   • Batch normalization
   • QLoRA
   • Tokenization
6. Why does per-channel or per-group scaling help quantized models?
   • It assigns separate scale factors to slices of a weight matrix to preserve precision
   • It removes the need for any weights
   • It makes the model larger
   • It converts text into audio
7. A four-bit model’s raw weights fit in available VRAM, but serving fails out of memory. What could explain this?
   • Every four-bit weight expands permanently to a gigabyte
   • Quantization guarantees no runtime allocations
   • Only the model’s filename consumes the remaining memory
   • Caches, scales, unquantized tensors, or workspace need additional memory
8. A team quantizes to INT8 but sees no inference speedup. What should it check?
   • Whether the training data are alphabetized
   • Whether the workload uses efficient supported kernels on its hardware
   • Whether lower bit width always guarantees faster execution
   • Whether the model must have changed languages