<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Yong Lai</title><description>Running AI systems in production: what works, what breaks, and why.</description><link>https://yonglai.github.io/</link><item><title>A passing /health check doesn&apos;t mean your vLLM engine can decode</title><link>https://yonglai.github.io/posts/health-check-passes-engine-frozen/</link><guid isPermaLink="true">https://yonglai.github.io/posts/health-check-passes-engine-frozen/</guid><description>I froze the engine of a single-GPU vLLM pod on k3s. The default health check stayed green and the pod kept receiving traffic until I thawed it 108 seconds later. A dummy-decode readiness probe took it out of rotation after 37.</description><pubDate>Sun, 08 Nov 2026 00:00:00 GMT</pubDate></item><item><title>What a 6-minute ML outage costs: 372 failed requests and 5.7% of the month&apos;s error budget</title><link>https://yonglai.github.io/posts/outage-error-budget/</link><guid isPermaLink="true">https://yonglai.github.io/posts/outage-error-budget/</guid><description>I scaled a model server to zero to test my monitoring. The alert fired after 2 minutes, 372 requests failed, and the latency graph went blank instead of red.</description><pubDate>Sat, 07 Nov 2026 00:00:00 GMT</pubDate></item><item><title>FlashInfer&apos;s version guard checked the compiler, not the headers</title><link>https://yonglai.github.io/posts/version-guard-checked-the-compiler/</link><guid isPermaLink="true">https://yonglai.github.io/posts/version-guard-checked-the-compiler/</guid><description>Why vLLM crashed at sampler warmup on my machine, how a FlashInfer guard picked a CUB API using the wrong version, and why the fix was deleting nine lines instead of repairing the guard.</description><pubDate>Tue, 03 Nov 2026 00:00:00 GMT</pubDate></item><item><title>Three wrong answers, three checks: Bedrock&apos;s grounding tiers against a claim-level verifier</title><link>https://yonglai.github.io/posts/three-wrong-answers-three-checks/</link><guid isPermaLink="true">https://yonglai.github.io/posts/three-wrong-answers-three-checks/</guid><description>Bedrock&apos;s default grounding check passed all three wrong answers from my insurance agent. The Standard tier caught one, a claim-level checker caught all three, and each had a cost. Injected errors made the two Bedrock grounding checks look better than they were.</description><pubDate>Mon, 02 Nov 2026 00:00:00 GMT</pubDate></item><item><title>Fusion, not CUDA graphs: where the batch-1 decode speedup actually comes from</title><link>https://yonglai.github.io/posts/fusion-not-cuda-graphs/</link><guid isPermaLink="true">https://yonglai.github.io/posts/fusion-not-cuda-graphs/</guid><description>On a small model at batch 1, kernel fusion delivered 95% of torch.compile&apos;s 2.87x decode speedup and CUDA graphs added the rest. Plus a benchmark prompt that nearly fooled me.</description><pubDate>Sat, 31 Oct 2026 00:00:00 GMT</pubDate></item></channel></rss>