Author

aiterms

What is the KV cache?

The KV cache, or key-value cache, is a memory buffer that stores past attention keys and values so a transformer can generate the next token without recomputing everything from scratch.

Why it matters

In autoregressive generation, each new token depends on all previous tokens. Without a KV cache, the model would repeatedly re-run attention over the entire prefix at every step, which is much slower than necessary.

You reach for a KV cache when you want faster text generation, lower latency, or higher throughput in decoder-only transformers and similar LLM serving setups. In practice, it is one of the main reasons incremental decoding is feasible at useful speeds.

How it works

A transformer attention layer turns each token into three things: queries, keys, and values. During generation, the model only needs to compute a new query for the latest token, but it still needs access to the keys and values for all previous tokens to attend over them.

The KV cache stores those previous keys and values layer by layer. When the next token is generated, the model appends the new token’s key and value to the cache, then uses the cached history plus the new query to compute attention. This avoids recomputing keys and values for the whole prefix on every step.

Most implementations keep a separate cache for each attention layer, and the cache grows as the sequence grows. That speedup comes with a memory cost: longer contexts and larger batch sizes require more GPU memory, so serving systems often balance cache size, precision, and batching strategy.

Tiny concrete example

Suppose a model has already processed:

“The cat sat”

It has cached the key/value tensors for those three tokens in each layer. Now it wants to generate the next token.

• Without a KV cache: it would recompute attention for “The cat sat” again, then add the new token.
• With a KV cache: it reuses the stored keys and values for “The cat sat,” computes attention for the new position, and produces the next token faster.

Common pitfalls / when NOT to use it

• It is not free memory. A longer conversation means a larger cache, which can become the bottleneck.
• It helps generation, not training. Training usually processes many tokens in parallel, so the incremental decoding trick is less relevant there.
• It does not make attention cheap in the limit. The model still attends over the growing history; the cache mainly removes redundant recomputation.
• It can complicate batching and serving. Systems need to manage per-request cache state carefully, especially with many concurrent generations.
• It is model-architecture dependent. The term is most common for transformer decoders; the exact implementation differs across libraries and serving stacks.

Related terms

• What is a token (and tokenization)?
• What is a transformer?
• What is attention (self-attention)?
• What is temperature in an LLM?
• What is top-p / nucleus sampling?
• What is a mixture-of-experts (MoE) model?
• What is a parameter (model size)?
• What is a hallucination?
• What is grounding?

同じ著者の記事

What is LLM-as-a-judge?

LLM-as-a-judge is the practice of using a large language model to evaluate or rank outputs from another model, system, or workflow, instead of relying only on human reviewers or fixed metrics. It solves a common problem: many AI outputs are hard to score with simple automated checks. If you are judging a summary for usefulness, a chatbot answer for helpfulness, or two candidate responses for preference, exact-match metrics often miss the point. Teams reach for LLM-as-a-judge when they need: **Fa

papoo.work

What is an LLM eval (evaluation)?

An LLM eval, or evaluation, is a test or set of tests used to measure how well a large language model behaves on a task or under a specific set of conditions. If you build with LLMs, you need a way to answer practical questions like: Does this prompt change improve answer quality? Is the model following instructions more reliably? Did a new model version get worse on safety, latency, or factuality? An eval gives you a repeatable way to compare models, prompts, retrieval setups, or fine-tunes. In

papoo.work

What is a hallucination?

A hallucination is when an AI system generates output that sounds plausible but is wrong, unsupported, or made up. Hallucinations matter because they break trust. If you use a model for search, summarization, coding help, or customer support, a confident but false answer can mislead users, create bugs, or spread bad information. In practice, teams care about hallucinations whenever they need factual accuracy, source grounding, or reliable behavior. For creative writing, they may be less of a pro

papoo.work

What is a parameter (model size)?

A parameter is a learned number inside a model, and “model size” usually means how many of those learned numbers the model has. Parameters are the part of a neural network that gets updated during training, so they determine how much capacity the model has to fit patterns in data. You’ll care about parameter count when you want to compare models, estimate training cost, understand memory use, or get a rough sense of how much hardware a model needs. In practice, “bigger” often means more capable,

papoo.work

What is a mixture-of-experts (MoE) model?

A mixture-of-experts (MoE) model is a neural network that routes each input, or part of an input, to only a small subset of specialized submodels (“experts”) instead of using the whole model every time. MoE models are useful when you want much larger capacity without paying the full compute cost on every token or example. In practice, that means you can scale model size to improve quality while keeping inference and training more efficient than a dense model of similar total parameter count.

papoo.work

What is a token (and tokenization)?

A token is a chunk of text that a language model reads and predicts, and tokenization is the process of splitting text into those chunks. LLMs do not operate directly on raw characters or words; they operate on tokens. That matters because: Input limits are measured in tokens, not characters. Cost and latency are often tied to token count in LLM APIs. Model quality depends on how text is tokenized: the same sentence can become very different token sequences across tokenizers.

papoo.work

What is a transformer?

A transformer is a neural network architecture that processes sequences by using attention to compare each token with the others, making it a flexible foundation for modern language models and many other sequence tasks. Before transformers, many sequence models had to read text step by step, which made long-range dependencies harder to handle and training slower to parallelize. Transformers solve this by letting the model look at many positions in the input at once. In practice, you reach for a

papoo.work

What is attention (self-attention)?

Attention, or self-attention, is a way for a model to decide which other words or tokens in the same input matter most when interpreting each token. Self-attention solves a basic language problem: the meaning of a word often depends on other words nearby or far away. It lets a model build a context-aware representation of each token instead of treating tokens mostly independently or relying only on fixed-size windows. In practice, this is why Transformers work so well for text, code, and other s

papoo.work

What is temperature in an LLM?

Temperature is a sampling setting that controls how random or deterministic an LLM’s next-word choices are when it generates text. Temperature is one of the simplest ways to trade off consistency versus variety. Use a lower temperature when you want stable, repeatable answers: classification-like outputs, structured data, summaries, or code. Use a higher temperature when you want more diversity: brainstorming, creative writing, alternate phrasings, or exploratory chat. In practice, many

papoo.work

What is top-p / nucleus sampling?

Top-p sampling, also called nucleus sampling, is a way to make a language model pick the next token from only the smallest set of likely options whose combined probability reaches a chosen threshold. If you always choose the single most likely next token, model outputs can become repetitive and brittle. If you sample from all tokens, outputs can get noisy and incoherent. Top-p gives you a middle ground: it keeps the model flexible, but trims away the long tail of very unlikely tokens before samp

papoo.work