llama.cpp cutoffs, cache flags, and PR text dispute

llama.cpp saw three low-volume threads on kernel cutoffs, model addition quality, and cache allocation. All were single-message items from the ggml-org repository.

sm_86 MMVQ cutoff requested

A feature request asks maintainers to add an sm_86 entry to the MMVQ cutoff table in llama.cpp. The requester reports that Q4_0 crosses at 7 on an A10 and estimates a 9.1 percent gain. Quantization and Ampere GPU users track these tables because they control which matrix-vector kernels run at inference time.

AI-generated text criticized in GLM PR

A comment on the pull request that adds GLM-5.3-Flash (glm5next) calls the accompanying text AI-generated slop that breaks project rules. Only one participant appears in the thread. Contributors who submit new models to llama.cpp face scrutiny over whether descriptions are written by hand.

Extra V-cache allocation questioned

An issue flags possible unnecessary V-cache allocation for DSA and the Lightning Indexer inside llama.cpp. The report is a single opening message with no further discussion yet. Implementers of those indexers care because surplus cache pressure affects memory footprint on long contexts.