Problemsby GlyphTrouble with vLLM cache
4 posts from 4 people4d activeRising, +2 this week
Posts per day
The posts behind it
[RFC]: Large-page KV cache allocator: one KV cache group per spec
...the details). A drafter can do the same (Kimi-K3 with DSpark: 4 groups become 20). **The workarounds add concepts that every component must understand.** `UniformTypeKVCacheSpecs` merges layers of one type with different sizes into one spec. `hash_block_size` and...
[Bug]: AOT compile cache never loads for Qwen3.5 models: "Kernel index 1 not found in id_to_kernel"
...defaults to 1 on torch >= 2.10). ### 🐛 Describe the bug On a warm start, vLLM fails to load the AOT-compiled artifact it saved on the previous start, and recompiles. **Command** **Steps** 1. Start once. The log ends compilation with `saved AOT compiled function to...
External speculators (DSpark Qwen3 draft, DFlash2) crash with kv_cache_dtype=nvfp4_ds_mla on GLM-5.3 NVFP4: draft-model code assumes split k/v
...kernel routing #58744 added, or #59342 (native nvfp4 sparse decode). Both drafters are unusable with NVFP4 KV on 0.31 stock + #58744.
[Bug]: Engine fails to start with `--enable-expert-parallel` and `--load-format ipc_cache`; the disk fallback fails too
### Your current environment Environment summary (trimmed from vllm collect-env ) ### 🐛 Describe the bug With expert parallelism enabled (EP size > 1), an engine started with `--load-format ipc_cache` crashes while it builds the model. The default `fallback=true` does not help:...
Companies and products named
About this problem
Evidence
4 posts from 4 people in 1 place, about 1 a week over 28 days. Mostly on vllm-project/vllm. Tools named alongside: GitHub, NVIDIA, Kernel.
Frustration Frustration 2 of 3· Seen on GitHub issues
How it was grouped
Posts about vLLM with no known issue, sharing the word "cache". First post Oct 2, 2026, latest Oct 29, 2026. Corroborated: 3 or more posts from 2 or more people or places. Method
Other vLLM problems
Running vLLM locally hits memory and speed limits
13 posts from 13 people on GitHub issues, Reddit.
vLLM upgrades break working setups
17 posts from 17 people on GitHub issues.
vLLM runs out of context on real work
5 posts from 5 people on GitHub issues.
vLLM vulnerabilities are hard to triage and patch
3 posts from 3 people on Glyph Ibis.
vLLM users keep filing bugs
Top post: "Qwen3.8-flash-next 0% MTP acceptance rate in disaggregated PD serving", one of 44 posts from 40 people on GitHub issues, Reddit.
vLLM GPU access and drivers are painful
5 posts from 3 people on GitHub issues.
Trouble with vLLM performance
Top post: "Waiting requests re-walk their whole prefix-cache hit on every scheduling step", one of 3 posts from 3 people on GitHub issues, Reddit.
RAG on vLLM returns irrelevant or missing context
3 posts from 3 people on GitHub issues.
Trouble with vLLM installation
Top post: "v0.30.0 CPU wheels now require manylinux_2_39 (glibc 2.39), breaking installation on Ubuntu 22.04/Rhel9", one of 3 posts from 3 people on GitHub issues.
Trouble with vLLM decoding
Top post: "Support Qwen3.x MTP (Multi-Token Prediction) for speculative decoding", one of 3 posts from 3 people on GitHub issues.
Fine-tuning vLLM is costly and unpredictable
3 posts from 3 people on GitHub issues.
vLLM containers are slow, bloated or crash
2 posts from 2 people on GitHub issues. Unverified.
History