AI tooling proposals in llama.cpp and Transformers

Proposals in core open-source AI libraries focused on extending fine-tuning support and trimming import overhead. Activity was limited to single-message feature requests in llama.cpp and Hugging Face Transformers.

Backward kernels for gated delta-net fine-tuning in llama.cpp

A feature request filed against ggml-org/llama.cpp asks for backward operations so hybrid gated delta-net models can be fine-tuned. The request covers kernels for GATED_DELTA_NET, CONCAT, SSM_CONV, L2_NORM, and SIGMOID, and references a fork that already implements the backward paths. Developers building or adapting these hybrid architectures inside llama.cpp need the missing gradients to move beyond inference-only use.

Lazy imports to lighten tokenizer loading in Transformers

A report in huggingface/transformers notes that merely loading a tokenizer pulls in torch and scikit-learn whenever those packages are installed, consuming roughly two seconds of a three-second import. The suggested fix is to switch the relevant paths to lazy imports. Users who load tokenizers in lightweight scripts or constrained environments would see faster startup without changing public APIs.