Quiz: QLoRA and Quantization
30 questions on QLoRA, 4-bit quantization, and memory-efficient training.
Question 1
What does QLoRA combine?
A) Quantization and full fine-tuning
B) 4-bit quantization and LoRA
C) 8-bit and 16-bit training
D) Multiple models
Show Answer
Correct Answer: B
QLoRA combines 4-bit quantization of the base model with LoRA adapters for efficient fine-tuning.
Question 2
What quantization format does QLoRA use by default?
A) INT8
B) FP4
C) NF4 (4-bit NormalFloat)
D) Binary
Show Answer
Correct Answer: C
QLoRA uses NF4 (4-bit NormalFloat), designed for normally distributed weights.
Question 3
How much VRAM does QLoRA need for a 7B model?
A) ~28GB
B) ~14GB
C) ~6GB
D) ~1GB
Show Answer
Correct Answer: C
QLoRA requires approximately 6GB VRAM for a 7B parameter model.
Question 4
What is double quantization?
A) Running quantization twice
B) Quantizing the quantization constants
C) Using two GPUs
D) Double precision training
Show Answer
Correct Answer: B
Double quantization quantizes the quantization constants themselves, saving additional memory.
Question 5
What library provides BitsAndBytesConfig?
A) peft
B) transformers (using bitsandbytes)
C) torch
D) accelerate
Show Answer
Correct Answer: B
BitsAndBytesConfig is provided through transformers, which uses the bitsandbytes library.
Questions 6-30
(Additional questions covering: prepare_model_for_kbit_training, gradient checkpointing, gradient accumulation, paged optimizers, NF4 vs FP4, memory calculations, compute dtype, optimization combinations, troubleshooting OOM errors, and practical implementation details)
Show All Remaining Answers
Q6-Q30 cover:
- prepare_model_for_kbit_training purpose (B: prepares quantized model for training)
- Gradient checkpointing trade-off (B: compute for memory)
- Effective batch size calculation (C: batch_size × accumulation_steps)
- Recommended compute dtype (C: bfloat16)
- Double quantization savings (B: ~11%)
- Paged optimizer purpose (B: pages to CPU on memory pressure)
- QLoRA quality vs full FT (B: ~98%)
- NF4 advantage (B: optimal for normally distributed weights)
- Memory reduction from QLoRA vs LoRA (C: ~60%)
- When to use QLoRA (B: limited GPU memory)
Results
| Score | Level |
|---|---|
| 27-30 | Excellent |
| 21-26 | Good |
| 15-20 | Average |
| Below 15 | Review needed |