Fine-tuning Qwen is costly and unpredictable

MLOps2 posts from 2 people2d active+22 posts in the last 7 days, 0 the 7 days before (new)Unverified

Posts per day2 posts, Sep 19 to Sep 25

The posts behind it

2, newest first

Post	Where	Date
...a small list, but that will be very org dependent. Programming is very open ended, I struggle to see the 'deterministic set of choices' here. I'm looking at Jev/Kev like models for some control over tool calling and similar, but I also think that fine-tuning out behaviors (like...Hacker News commentsverdverm	Hacker News commentsverdverm	Sep 223 days ago
[Bug]: Decode throughput drops ~3.3x from 0.26.0 to 0.29.0 on H100 (Qwen3.6-35B-A3B-FP8, same config, same backends)...`ignore_eos`, ~122-token prompt, median of 5 repetitions (single measurements are unusable, see below): \| \| c=1 \| c=12 \| inter-token latency \| \|---\|---\|---\|---\| \| A (0.24.0) \| 138.8 tok/s \| 1183.3 tok/s \| 7.08 ms \| \| B (0.26.0) \| 148.3 tok/s \| 1295.0 tok/s \|...vllm-project/vllmyeighta	vllm-project/vllmyeighta1 reactions, 5 replies	Sep 196 days ago

Companies and products named

Company or product	Posts naming it
Qwen	2
vLLM	1
PyTorch	1
NVIDIA	1

About this problem

Evidence

2 posts from 2 people in 2 places, about 2 a week over 4 days. Mostly on Hacker News comments, vllm-project/vllm. Tools named alongside: vLLM, PyTorch, NVIDIA.

Frustration Frustration 3 of 3· Seen on Hacker News, GitHub issues

How it was grouped

Posts that state a pain and match the "fine-tuning" rule about Qwen. First post Sep 19, 2026, latest Sep 22, 2026. Unverified: fewer than 3 posts, or one person or place only. Method

The same problem elsewhere

Posts per day

The posts behind it

Companies and products named

Rising problems by email