LLM Serving
14 practice items · needed by 3 roles
Easy 1Medium 3Hard 10
Practice questions
- Memory for 4-bit weightsmedium
- KV cache size for one sequencehard
- Speculative decoding and output qualityhard
- Why serving engines page the KV cachehard
- Daily cost of an LLM featurehard
- Why long contexts eat GPU memoryhard
- What streaming improveshard
- Memory for 7B parameters in fp16easy
- KV cache size for one long requestmedium
- Serving many fine-tuned variantsmedium
- Why LLM servers use continuous batchinghard
- Where to use tensor parallelismhard
- KV cache with grouped-query attentionhard
- Semantic cache returns wrong answershard