# NVIDIA's TensorRT-LLM more than doubles DeepSeek cache capacity in one test

> Source: <https://runtimewire.com/article/tensorrt-llm-deepseek-fp4-cache-1-90m-tokens>
> Published: 2026-11-16 15:33:51+00:00

# NVIDIA's TensorRT-LLM more than doubles DeepSeek cache capacity in one test

**NVIDIA's batch-based allocator gave MLA layers 38.9 GiB of a 43 GiB DeepSeek-R1 cache reservation, up from 16.1 GiB in one reported layout.**

        By [RuntimeWire Staff](https://runtimewire.com/author/runtimewire-staff)
        · Published 

NVIDIA merged pull request #21313 into TensorRT-LLM main on November 16th, 2026, changing how the dense NVFP4 multi-head latent attention cache is sized and reused for DeepSeek-V3/R1-style models. In contributor Qi Zhang's DeepSeek-R1 serving run with multi-token prediction, a context worker with 43 GiB reserved for KV cache previously gave its MLA layers 16.1 GiB, enough for 787,000 tokens. The new allocator gave those layers 38.9 GiB and 1.90 million tokens. The pull request is credited to the GitHub account Tracin, and commit signatures in the merged history identify Qi Zhang.…
