llama.cpp quantization pushback and related proposals

llama.cpp drew the bulk of attention with a maintainer dispute over new quantization types plus several feature and testing threads, while Transformers saw one narrow validation proposal. Activity stayed technical and centered on model loading, low-bit formats, and server compatibility.

Maintainer objects to IQ2_NL and IQ3_NL types

A pull request seeking to add IQ2_NL and IQ3_NL quantization types for CPU use in llama.cpp met resistance from a maintainer. The objection cites maintenance burden and the presence of ML-generated code. Developers tracking quantization expansions should note the reluctance to accept these additions.

OpenAI Responses API support requested for llama.cpp server

A lone feature request asks the llama.cpp server to support the OpenAI Responses API endpoint. The proposal seeks parity with that interface for existing server users. Compatibility with newer OpenAI surfaces remains the practical interest for readers running local servers.

Optional checkpoint validation proposed in Transformers

A single opening message in the Transformers repository proposes making checkpoint validation optional inside from_pretrained, including stricter key checks. The change would let callers control how rigidly keys are validated on load. Model-loading workflows that need flexible or strict behavior are the audience for the idea.

Escha-W2 2-bit format support requested

Another feature request calls for llama.cpp to add the Escha-W2 2-bit quantization format. The message stands alone with no further discussion yet. Readers watching very-low-bit options have a new candidate format under consideration.

Prompt-processing speedup PR draws mixed test results

Users exercised pull request 28623 in llama.cpp, aimed at faster prompt processing, and reported RAM pressure plus problems on two-GPU setups. Three participants exchanged notes on the observed behavior. The feedback flags concrete limits for multi-GPU and memory-constrained configurations.