<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <title>Turingfold AI: Inference Engineering Research</title>
  <link href="https://www.turingfold.ai/"/>
  <link rel="self" href="https://www.turingfold.ai/feed.xml"/>
  <id>https://www.turingfold.ai/</id>
  <updated>2026-11-16T00:00:00Z</updated>
  <author><name>turingfold</name></author>
  <entry>
    <title>vLLM expands encoder-disaggregated serving.</title>
    <link href="https://www.turingfold.ai/brief/encoder-disaggregation/"/>
    <id>https://www.turingfold.ai/brief/encoder-disaggregation/</id>
    <published>2026-11-16T00:00:00Z</published>
    <updated>2026-11-16T00:00:00Z</updated>
    <category term="Daily Brief"/>
    <summary>Metadata-only requests, remote embeddings, and shared encoder outputs extend the encoder / prefill / decode path.</summary>
  </entry>
  <entry>
    <title>FlashInfer 0.7.4 adds batch-invariant kernels.</title>
    <link href="https://www.turingfold.ai/brief/flashinfer-batch-invariance/"/>
    <id>https://www.turingfold.ai/brief/flashinfer-batch-invariance/</id>
    <published>2026-11-16T00:00:00Z</published>
    <updated>2026-11-16T00:00:00Z</updated>
    <category term="Daily Brief"/>
    <summary>Selected GEMMs and quantizers can produce the same output-row bits across batch sizes on B200, B300, and Rubin GPUs.</summary>
  </entry>
  <entry>
    <title>vLLM adopts the LibTorch Stable ABI for CUDA wheels.</title>
    <link href="https://www.turingfold.ai/brief/vllm-stable-abi/"/>
    <id>https://www.turingfold.ai/brief/vllm-stable-abi/</id>
    <published>2026-11-16T00:00:00Z</published>
    <updated>2026-11-16T00:00:00Z</updated>
    <category term="Daily Brief"/>
    <summary>From v0.33.0, compatible PyTorch upgrades can avoid rebuilding the vLLM CUDA package.</summary>
  </entry>
  <entry>
    <title>Prefill and decode want different things.</title>
    <link href="https://www.turingfold.ai/brief/prefill-decode/"/>
    <id>https://www.turingfold.ai/brief/prefill-decode/</id>
    <published>2026-11-16T00:00:00Z</published>
    <updated>2026-11-16T00:00:00Z</updated>
    <category term="Daily Brief"/>
    <summary>A literature note on independent scaling and the cost of moving KV state.</summary>
  </entry>
  <entry>
    <title>The fastest kernel depends on how you measure.</title>
    <link href="https://www.turingfold.ai/brief/kernel-measurement/"/>
    <id>https://www.turingfold.ai/brief/kernel-measurement/</id>
    <published>2026-11-16T00:00:00Z</published>
    <updated>2026-11-16T00:00:00Z</updated>
    <category term="Daily Brief"/>
    <summary>A literature note on eager execution, graph replay, and tuning decisions.</summary>
  </entry>
  <entry>
    <title>A serving system is more than a tokens/s number.</title>
    <link href="https://www.turingfold.ai/brief/serving-frontier/"/>
    <id>https://www.turingfold.ai/brief/serving-frontier/</id>
    <published>2026-11-16T00:00:00Z</published>
    <updated>2026-11-16T00:00:00Z</updated>
    <category term="Daily Brief"/>
    <summary>A literature note on total throughput, per-user speed, and latency constraints.</summary>
  </entry>
</feed>
