Data and AI struggles

2,114 problems · 377 companies

MLOps

Trouble with vLLM start

2 posts from 2 people2d activeFading, +0 this weekunverified

Posts per day

Posts per day2 posts, Oct 9 to Nov 21

The posts behind it

GitHub issuesvllm-project/vllm · liranschourNov 11

[Bug][KV Offload]: canonical_layout refuses to start with FlashInfer on pre-Blackwell GPUs: kernel-split blocks fail the canonical shape check

...**Scope.** The failing check is on shape, and kernel block selection does not depend on dtype. From reading the code, bf16 with FlashInfer should fail the same way; I have not run that case. Affected setups are FlashInfer + canonical_layout + block_size > 64 on any GPU that is...

GitHub issuesvllm-project/vllm · michael-gpuOct 9

[Bug]: token_indices_to_sample underflows past the request's query start with PP > 1 spec decode

### 🐛 Describe the bug `eagle_prepare_inputs_padded_kernel` computes `index_to_sample = q_last_tok_idx - num_rejected_tokens` without a lower bound. When a request's query carries no bonus-token row - which happens with pipeline parallelism, where the scheduler can schedule...

Companies and products named

2 posts
2 posts
1 post
1 post
About this problem

Evidence

2 posts from 2 people in 1 place, about 1 a week over 34 days. Mostly on vllm-project/vllm. Tools named alongside: NVIDIA, Kernel, GitHub.

Frustration Frustration 0 of 3· Seen on GitHub issues

How it was grouped

Posts about vLLM with no known issue, sharing the word "start". First post Oct 9, 2026, latest Nov 11, 2026. Unverified: fewer than 3 posts, or one person or place only. Method

Other vLLM problems

vLLM users keep filing bugs

Top post: "Qwen3.8-flash-next 0% MTP acceptance rate in disaggregated PD serving", one of 37 posts from 34 people on GitHub issues, Reddit.

Running vLLM locally hits memory and speed limits

27 posts from 26 people on GitHub issues, Reddit.

vLLM upgrades break working setups

21 posts from 21 people on GitHub issues.

vLLM vulnerabilities are hard to triage and patch

9 posts from 9 people on Glyph Ibis.

vLLM runs out of context on real work

7 posts from 7 people on GitHub issues.

Trouble with vLLM cache

Top post: "AOT compile cache never loads for Qwen3.5 models: Kernel index 1 not found in id_to_kernel", one of 7 posts from 6 people on GitHub issues.

Fine-tuning vLLM is costly and unpredictable

7 posts from 7 people on GitHub issues.

vLLM GPU access and drivers are painful

5 posts from 3 people on GitHub issues.

vLLM fails to return valid structured output

5 posts from 5 people on GitHub issues.

Trouble with vLLM serve

Top post: "gpt-oss can't be served through `vllm serve` with any tool/reasoning-parser flags", one of 4 posts from 4 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.

vLLM lock-in makes switching costly

3 posts from 3 people on GitHub issues.

History

Momentum went from steady to fading.

Nov 18, 2026

First seen with 2 posts; the top post now: "canonical_layout refuses to start with FlashInfer on pre-Blackwell GPUs: kernel-split blocks fail the".

Nov 17, 2026