LLM Interview Prep
Architecture & Theory
1. "Explain the 'KV Cache' and why it's critical for real-time LLM applications."
In autoregressive models, we predict tokens one by one. Each new prediction depends on the hidden states of all previous tokens. Instead of re-calculating the Key ($K$) and Value ($V$) dot products for the entire history every time, we store (cache) them. This reduces the time complexity for a single token generation from $O(L)$ to $O(1)$ relative to history length, significantly reducing latency.
2. "Why do we divide by $\sqrt{d_k}$ in the scaled dot-product attention formula?"
As the dimension $d_k$ grows, the magnitude of the dot product $QK^T$ increases. This results in very large values being passed to the Softmax function, pushing it into regions where gradients are nearly zero (vanishing gradients). Scaling by $\sqrt{d_k}$ keeps the variance of the dot product around 1, ensuring stable gradients during training.
3. "Compare LoRA and Prefix-Tuning as PEFT methods."
LoRA (Low-Rank Adaptation) adds trainable low-rank matrices to the existing weights (usually $Q$ and $V$). It doesn't increase the sequence length. Prefix-Tuning prepends trainable vectors (prefixes) to the keys and values of every layer. This effectively reduces the context window available for the actual user prompt, whereas LoRA is more architectural and "invisible" to the context window.
Training & Production
4. "How do you detect and mitigate 'catastrophic forgetting' when fine-tuning an LLM?"
Detect: Evaluate the model on a "general" benchmark (like MMLU) after fine-tuning on a specific task; if the general score drops significantly, forgetting has occurred. Mitigate: 1. Use a very low learning rate. 2. Use PEFT (LoRA) which freezes the base weights. 3. Use Experience Replay (mixing some general pre-training data into the fine-tuning set).
5. "What is 'RLHF' and why can't we just use Supervised Fine-Tuning (SFT) for alignment?"
SFT helps the model learn the structure of an answer, but it's limited by the "average" quality of the human labelers. RLHF allows the model to explore multiple responses and learns from a Reward Model that captures complex human preferences (e.g., tone, safety, nuance) that are hard to explicitly write as "perfect" labels in SFT.
6. "Describe the 'Chinchilla' scaling laws in your own words."
It suggests that for a given amount of compute, most early LLMs were too large and under-trained. The optimal approach is to scale the number of parameters and the number of training tokens equally. Specifically, for every parameter, you should train on approximately 20 tokens.
Creative & Problem Solving
7. "An LLM is hallucinating a legal fact. How do you solve this without training?"
Use RAG (Retrieval-Augmented Generation). 1. Retrieve the actual law text from a verified database. 2. Inject that text into the prompt context. 3. Instruct the model to "Answer only using the provided text." This anchors the model in reality and provides a "paper trail" via citations.
8. "Your model is too large to fit on a single 80GB A100 GPU. What are your options?"
Quantization: Use 4-bit or 8-bit precision (reduces size by 2-4x). 2. Model Parallelism: Split the layers or heads across multiple GPUs. 3. PEFT (LoRA): Train only a tiny fraction of weights whilekeeping the rest on CPU/Disk (if training). 4. Offloading: Move part of the model to System RAM (slower).
Last updated