Problemsby IbisFine-tuning vLLM is costly and unpredictable
MLOps3 posts from 3 people3d active+11 post in the last 7 days, 1 the 7 days before (steady)
Posts per day
The posts behind it
3, newest first[Bug]: dtype="float32" + enable_lora crashes LFM2 startup — lora_expand Triton op asserts `weight.dtype in [torch.float16, torch.bfloat16]`### Your current environment ### 🐛 Describe the bug ## Description With `enable_lora=True`, setting `dtype="float32"` (a documented, user-settable value) makes the engine die during the startup dummy forward, inside the LoRA expand Triton op: Nothing validates the pair...vllm-project/vllm · jyfct3564 repliesOpen ↗Discussion: configurable decoder loops with shared parameters and independent execution cachesI would like to discuss how Transformers should represent configurable decoder loops. For experiments with pretrained decoder-only models, it is useful to load the original checkpoint, repeat a layer or a range of layers, share its parameters, and keep separate state histories...huggingface/transformers · bebetterest1 reactionsOpen ↗[Bug]: PEFT sequence-classification LoRA adapters fail to load on RoBERTa / XLM-RoBERTa...but adapters trained with PEFT `task_type="SEQ_CLS"` still fail to load: PEFT saves the classification head with these adapters (`modules_to_save: ["classifier", "score"]`). `parse_fine_tuned_lora_name` only accepts a saved head whose tensors sit directly under `classifier` or...vllm-project/vllm · subin97 repliesOpen ↗
Companies and products named
About this problem
Evidence
3 posts from 3 people in 2 places, about 2 a week over 10 days. Mostly on vllm-project/vllm, huggingface/transformers. Tools named alongside: NVIDIA, Hugging Face, GitHub, PyTorch.
Frustration Frustration 0 of 3· Seen on GitHub issues
How it was grouped
Posts that state a pain and match the "fine-tuning" rule about vLLM. First post Oct 6, 2026, latest Oct 15, 2026. Corroborated: 3 or more posts from 2 or more people or places. Method
The same problem elsewhere
Fine-tuning Hugging Face is costly and unpredictableFine-tuning is costly and results are unpredictableFine-tuning Qwen is costly and unpredictableFine-tuning llama.cpp is costly and unpredictableFine-tuning OpenAI is costly and unpredictableFine-tuning GitHub is costly and unpredictableFine-tuning Claude Code is costly and unpredictableFine-tuning Gemini is costly and unpredictableFine-tuning PyTorch is costly and unpredictableFine-tuning Claude is costly and unpredictableFine-tuning NVIDIA is costly and unpredictableFine-tuning ChatGPT is costly and unpredictable
Other vLLM problems
vLLM upgrades break working setupsTrouble with vLLM decodingRunning vLLM locally hits memory and speed limitsvLLM GPU access and drivers are painfulRAG on vLLM returns irrelevant or missing contextTrouble with vLLM installationvLLM runs out of context on real workTrouble with vLLM performancevLLM runs out of memoryvLLM lock-in makes switching costlyvLLM users keep filing bugs
History
unverified to corroborated
fading to steady
steady to fading
new to steady
Fine-tuning vLLM is costly and unpredictable (2 posts)