Repository Issue Activity (beta)

#2990 eleuther_eval recipe: unpinned GPT-2 fetch-and-discard; multimodal path silently scores empty answers; generation budget hardcoded to 256 ignoring max_gen_toks

#2989 [Bug] Torchtune recipes crash on Apple Silicon (MPS) due to hardcoded CUDA/CPU memory and device assumptions

#1960 Enable "apply_lora_to_output" in models with tied embedding

#2984 Add support to IBM Granite models

#2943 Bug in `async_grpo_full_finetune` recipe: TrainingWorker receives identical advantages

#2700 Log config to output only on rank zero

#2478 Add add_end_token to the Gemma Tokenizer

#2971 Arbitrary code execution via the _component_ field in an untrusted recipe config

#2968 Backport workflow-hardening fix (unpinned-uses) to 5 release branches

#2962 [Proposal] Native support for SCAO (2nd-Order Optimizer) for memory-bound QLoRA recipes

#2012 Add custom dataset section in Basics in live docs

#2883 [IMPORTANT] The future of torchtune

#2890 torchtune v0.6.1 fresh install defaults to pyarrow v21.0.0 causing error

#2954 Feature : (Diagnostic) Automated FSDP Compute–Communication Overlap Diagnostic in Profiler

#2928 Request to implement SCALE optimizer

#2931 lr_scheduler bug

#2907 DistributedCheckpointer argument conflicts

#2917 H2D takes large time overhead

#2946 kagglehub changed function: set_kaggle_credentials renamed to set_kaggle_api_token

#2930 Error while full finetuning Llama 3.2 Vision

#2939 When will the qwen3-vl model be supported?

#2924 Qwen 3 0.6B full_finetune_distributed breaks with context parallelism enabled

#2925 support packing with flash_attn v2?

#2923 qwen3 tokenizer think token

#2929 can you add support for LlamaGuard (meta-llama/Llama-Guard-3-8B)?