Qlora

Match each 4-bit format to its niche in 2026 inference and fine-tuning.

Merging a LoRA back into an NF4-quantized base hits one subtle problem: name it and explain why

Adam stores two fp32 buffers per trainable parameter: what does that imply for LoRA vs full FT on a 7B?

QLoRA OOMs at step 7000 after one clean epoch on a 24 GB GPU: which non-obvious causes deserve a look?

Match each PEFT method to the standout advantage that defines it.

Paged optimizer in QLoRA, what gets paged, and to where?

FP4 vs NF4, pick the answer that captures the actual structural difference

Explain what gets quantized the second time in QLoRA's double-quantization step

What does bnb_4bit_compute_dtype actually control in a QLoRA config?

How does QLoRA fit a 65B fine-tune onto a single 48GB GPU?

Which statements about QLoRA are true?

Why does QLoRA use NF4 instead of uniform INT4?

Match PEFT method to the scenario it fits best

Misconception: 'LoRA is just a quantization technique'

Match training regime to sensible peak LR range

Estimate peak VRAM for a 7B full-FT run in bf16 + Adam

LoRA vs full FT on 7B + 10k examples: order of magnitude cost comparison

What does LoRA do, and why is it popular for fine-tuning?