$ the-wire · showcase
TRL reads tokenizers for text-only models, chat-ui queues messages mid-run
By RepoJournal · Filed · About Hugging Face · Composed from the cited sources · methodology
TRL fixes the trainer path that forced a tokenizer on anyone loading a multimodal checkpoint as a plain causal LM, chat-ui makes ML Intern interruptible, and transformers both deprecates two old input names and gains a new TTS model.
Load the tokenizer for models that take no images huggingface/trl
When no processing_class was passed, the trainers loaded AutoProcessor, so a Qwen3.5 checkpoint loaded with AutoModelForCausalLM went down the VLM path and users had to pass the tokenizer by hand. SFT, DPO, KTO, GRPO, RLOO and Distillation now check the new model_accepts_images flag and load AutoTokenizer when the model takes no images, reading input_modalities from transformers 5.21 as the sig...
Add support for vLLM 0.32.0 huggingface/trl
vLLM 0.32.0 changed /inference/v1/generate logprobs from OpenAI-style "token_id:<id>" placeholder strings to a typed shape with integer token IDs, and VLLMClient.generate parses that endpoint. Without the update, multimodal server-mode training on the --enable-scale-out path raises KeyError on the first batch that requests logprobs.
add btw side questions for ml intern huggingface/chat-ui
A message starting with "/btw " in an ML Intern conversation is not sent to the agent; the answer streams into a small dismissible panel above the composer, visible only to you and never stored. It works while a turn is running, while parked on a wait, or after it finished. A companion change queues ordinary messages typed mid-turn and delivers them at the next tool step, or cuts a wait short a...
Deprecate old input names (#49424) huggingface/transformers
transformers deprecated two old input names, keeping the new spellings as the supported path while the old ones warn; video_llava was fixed up to match. Code that still passes the deprecated kwargs through the processor will work for now but is on notice.
Document training on AMD GPUs huggingface/trl
A new "Training on AMD GPUs" docs page sits next to the Trainium one and spells out where ROCm diverges from CUDA: install PyTorch's ROCm build, use the AITER attention kernel because flash-attn2 has no ROCm build, which drops Gemma 2's attention soft-capping, and supply a separate vLLM ROCm build, since vLLM is not yet in CI.
Add torch-stable-abi-shim huggingface/kernels
Kernels vendors the header-only Torch API into a shim, so a kernel that needs newer headers can build on older Torch versions. Stable ABI kernels are built against the newest Torch a given backend supports, which for CUDA 12.8 means Torch 2.11, and the shim covers the header gap from there.
Action items
