Nick Barcet

A tricky LLM interview question: You're serving a reasoning model on vLLM, and it keeps running out of GPU memory on long traces. So you add KV cache compression and evict 90% of the cached tokens. VRAM usage stays as is and GPU still runs out of memory. Why? (answer below) Evicting 90% of the KV cache can free almost none of the memory it was using. This sounds counterintuitive, but it follows directly from how production servers store the cache today. The KV cache grows with every token a model generates. Each token appends its key and value vectors across every layer, and nothing is freed while generation continues. This is the dominant memory cost for reasoning models. If a 32K-token CoT caches ~32K tokens of KV vectors, a Qwen3-32B with 4-bit weights will run out-of-memory around 24K tokens on a 24GB GPU. One obvious solution is to keep the important tokens and drop the rest, since attention is sparse enough to allow it. But this does not solve the memory problem yet. The reason is paged attention, which is the memory manager behind vLLM and most production servers. Under the hood, it splits GPU memory into fixed physical blocks, each one holds the KV for about 16 tokens. This block returns to the allocator only when every slot inside it is empty. Since the eviction logic selects tokens by importance, and such tokens are scattered across blocks... ...so despite eviction, almost every block is left with at least some survivor tokens. For instance, if the logic evicts 14k of 16k tokens across 1,000 blocks, most likely every block will still have a token. This means the allocator frees almost nothing. Placing the new tokens into those freed slots is not ideal because it breaks the cache's layout. Say token 16,001 arrives, and it's placed in the slot the 40th token used to hold. The cache now reads position 38, then 16,001, then 41, so the cache is no longer in token order. Attention can still compute the right answer from that, but only if every slot now carries a separate note recording which position it actually holds. This introduces another bookkeeping cost that an in-order layout inherently avoids. So the cache is logically 90% smaller and still physically the same size. Many compression results miss this because they measure on pre-allocated contiguous tensors rather than a paged server. There's another problem. Eviction methods pick which tokens to keep by looking at the attention scores themselves (as expected). But fast attention kernels used in production, like FlashAttention, never save those scores. They compute attention in small pieces and throw the full score grid away as they go, which is also why they're fast. So the exact signal eviction methods need isn't available in memory. The workaround is to fall back to eager attention and build the full matrix, which gives up the speed FlashAttention was there to provide. NVIDIA published a method called TriAttention to solve both these problems. It never needs attention scores. Instead, it scores tokens from the geometry of the model's key and query vectors before RoPE is applied, where those vectors sit in stable clusters. For the memory problem, it runs a compaction pass every 128 decoded tokens. The surviving tokens slide forward to close the holes eviction creates, so whole blocks empty out and return to the allocator while the cache stays in token order. On long reasoning traces, the approach matches full-attention accuracy while decoding 2.5x faster and using 10.7x less KV memory. KV cache compression is a big infrastructure problem. The number that decides whether it works is the count of freed blocks, not the count of evicted tokens. You can find the NVIDIA write-up here: research.nvidia.com/labs/eai… I wrote a first-principles breakdown of how the KV cache works. It walks through why the model stores keys and values at all, why the cache grows with every token, and a comparison of LLM generation speed with and without KV caching. Read it below.

90% of your KV cache never gets reused. (prompt caching was never meant to fix it) if your system prompt and tool definitions are stable, prompt caching is the single highest-leverage optimization available today. cached input tokens get up to 90% cheaper, and hit rates of 60 to 85% are realistic. but it comes with one rigid rule. the cached portion must be an exact, byte-for-byte prefix of the new request. change a single character in that region and you get a full cache miss. that rule breaks in three situations you hit constantly: → RAG with multiple documents. you cached document A alone and document B alone. a query now needs both. document B's cached state was computed without any awareness of A, so it's invalid and gets recomputed from scratch. → document order changes. the same three documents appear in a different order across requests. every permutation is a cache miss, even though the content is identical. → growing conversation history. each new turn changes everything after the stable prefix, so earlier cached states beyond it become useless. Alibaba Cloud's production data shows how bad this gets: 10% of KV cache blocks serve 77% of all hits. the rest sits in storage and never gets reused, because prefix matching won't allow it. CacheBlend, a research paper from the LMCache team (EuroSys 2025 Best Paper Award), attacks exactly this. the insight is that in modern transformers, tokens overwhelmingly attend to their own local context. only a small fraction of tokens carry real connections across document boundaries. so instead of recomputing everything after the first cached document, CacheBlend reuses every document's cache as-is and selectively recomputes just those few boundary tokens. those are the small orange fixes between documents in the diagram, and they are the entire cost. the result is 2 to 4x faster processing on multi-document queries with no quality loss. the order problem disappears with it. shuffle the same documents however you like, and every permutation stays cached, where prefix caching recomputes all of them every time. the bottom of the diagram shows that side by side. that's the real shift: from caching prefixes to caching knowledge. every document in your knowledge base becomes a reusable cached asset, regardless of what order it appears in or what sits next to it. CacheBlend ships inside LMCache, the open-source cache management layer that runs outside the inference engine and integrates with vLLM, SGLang, and TensorRT-LLM, on both NVIDIA and AMD GPUs. check it out on GitHub: github.com/LMCache/LMCache (don't forget to star 🌟) i wrote the full breakdown of the architecture, including why cache management should never live inside your inference engine. the article is quoted below. stay tuned for more on this!