Skip to content
Problems.

vLLM runs out of context on real work

LLM apps2 posts from 2 people2d active+11 post in the last 7 days, 0 the 7 days before (steady)Unverified

Posts per day

Posts per day2 posts, Aug 28 to Oct 4

The posts behind it

2, newest first
PostDate
[RFC]: Octave KV, a native 3-bit KV cache for AMD GPUsPR: https://github.com/vllm-project/vllm/pull/59774 ### Proposal to improve performance ## Where this comes from We serve Qwen3.8-27B (GPTQ W4A16) on boxes with four RX 7900 XTX, 96 GB in total, with tensor parallel 4 and MTP. Our users send long prompts: whole repositories,...vllm-project/vllmJartXOct 22 days ago
[Bug]: GLM-5.3-Flash — ModelOpt NVFP4 checkpoints emit invalid UTF-8 byte tokens on SM120, while a compressed-tensors NVFP4 of the same model is clean### Your current environment Environment ### 🐛 Describe the bug Two ModelOpt-produced NVFP4 conversions of GLM-5.3-Flash emit **invalid UTF-8 byte-token sequences** during generation on SM120. A compressed-tensors NVFP4 conversion of the *same* model, served from the same image...vllm-project/vllmshing100Aug 285 weeks ago

Companies and products named

Company or productPosts naming it
vLLM2
NVIDIA2
GitHub1
About this problem

Evidence

2 posts from 2 people in 1 place, about 1 a week over 36 days. Mostly on vllm-project/vllm. Tools named alongside: NVIDIA, GitHub.

Frustration Frustration 3 of 3· Seen on GitHub issues

How it was grouped

Posts that state a pain and match the "context-window" rule about vLLM. First post Aug 28, 2026, latest Oct 2, 2026. Unverified: fewer than 3 posts, or one person or place only. Method

History

  • 2026-10-04 Added: vLLM runs out of context on real work (2 posts)

Rising problems by email

Mondays: the problems in data, tech and AI that grew fastest that week.

Double opt-in. Unsubscribe any time.