Problemsby GlyphFine-tuning Qwen is costly and unpredictable
4 posts from 4 people4d activeFading, +0 this week
Posts per day
The posts behind it
Stepping away from Benchmarks and Code, what models are you using for Creating writing projects
...mixed up but my creative writing harness takes care of this.. Qwen 3.8 27b finetunes, Disappointing, I guess the base model is really fucking bad for creative writing purposes and even with finetuning you can only do so much. It does follow instructions quite well but its...
tiny-aya models are loaded as CohereTokenizer in v5, and digits are split differently from tokenizer.json
...several model_types to MODELS_WITH_INCORRECT_HUB_TOKENIZER_CLASS. However, this approach does not work for Aya, because the class registered for cohere2 in transformers is CohereTokenizer itself, and there is no other Aya-specific class to switch to. One possible fix is to add...
...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...
[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** |...
Companies and products named
About this problem
Evidence
4 posts from 4 people in 4 places, about 1 a week over 21 days. Mostly on Hacker News comments, vllm-project/vllm, huggingface/transformers. Tools named alongside: vLLM, PyTorch, NVIDIA, Hugging Face.
Frustration Frustration 3 of 3· Seen on GitHub issues, Hacker News, Reddit
How it was grouped
Posts that state a pain and match the "fine-tuning" rule about Qwen. First post Sep 19, 2026, latest Oct 9, 2026. Corroborated: 3 or more posts from 2 or more people or places. Method
The same problem elsewhere
Fine-tuning is costly and results are unpredictable
12 posts from 12 people on Hacker News, Reddit, GitHub issues.
Fine-tuning Hugging Face is costly and unpredictable
9 posts from 9 people on GitHub issues, Reddit.
Fine-tuning vLLM is costly and unpredictable
4 posts from 4 people on GitHub issues.
Fine-tuning llama.cpp is costly and unpredictable
4 posts from 4 people on GitHub issues, Hacker News.
Fine-tuning Claude is costly and unpredictable
3 posts from 3 people on Reddit, GitHub issues.
Fine-tuning GitHub is costly and unpredictable
3 posts from 3 people on Reddit, GitHub issues.
Fine-tuning PyTorch is costly and unpredictable
3 posts from 3 people on GitHub issues, Reddit.
Fine-tuning NVIDIA is costly and unpredictable
3 posts from 3 people on GitHub issues.
Fine-tuning OpenAI is costly and unpredictable
3 posts from 3 people on GitHub issues, Hacker News.
Fine-tuning Gemini is costly and unpredictable
2 posts from 2 people on Reddit, GitHub issues. Unverified.
Fine-tuning ChatGPT is costly and unpredictable
2 posts from 2 people on Reddit, Hacker News. Unverified.
Fine-tuning Claude Code is costly and unpredictable
2 posts from 2 people on Reddit. Unverified.
Other Qwen problems
Running Qwen locally hits memory and speed limits
18 posts from 18 people on GitHub issues, Reddit, Stack Overflow.
Qwen users keep filing bugs
Top post: "Get ready for qwen4", one of 19 posts from 19 people on GitHub issues, Reddit, Community forums.
Qwen upgrades break working setups
5 posts from 5 people on GitHub issues.
Trouble with Qwen local
Top post: "Feature request: Local image-generation model support with hardware-tuned inference", one of 3 posts from 3 people on Reddit, Stack Overflow, GitHub issues.
Qwen rate limits and usage caps block work
2 posts from 2 people on Reddit, GitHub issues. Unverified.
Qwen runs out of context on real work
2 posts from 2 people on Reddit, GitHub issues. Unverified.
Qwen fails to return valid structured output
2 posts from 2 people on GitHub issues. Unverified.
History