llama.cpp GGUF bugs and model support, PyTorch Triton cache

Security reports against the GGUF loader in llama.cpp led the day, with parallel work on new model support and batching. PyTorch and Transformers saw smaller precompilation and MoE dispatch changes.

Unvalidated gguf_type read from untrusted GGUF

A report against ggml-org/llama.cpp states that the GGUF reader loads an out-of-range gguf_type from an untrusted file before validating it, at ggml/src/gguf.cpp:576. The path involves a forged key length. Anyone loading external GGUF models should treat the parser as not yet hardened against this class of input.

GLM-5.3-Flash support lands in llama.cpp

A pull request adds GLM-5.3-Flash (GLM5-Next) support to llama.cpp. Discussion covered a softmax crash fix, KV pooling changes, and a data race in a dummy model. The change widens the set of models the inference stack can run without external conversion steps.

Batches that mix embeddings and raw tokens

A pull request extends llama.cpp so a single batch can carry both embeddings and raw tokens, intended to speed multimodal prefill. The same change reports and fixes a scheduler reallocation bug. Multimodal pipelines that previously split those stages gain a combined path.

Hydrating Triton JIT cache from verified bundles

A PyTorch pull request introduces import_runtime_cache to seed Triton's JIT cache from a verified bundle. Review focused on test robustness and staging or concurrency edge cases. Precompilation workflows that need a trusted cache bootstrap are the direct audience.

Integer overflows in GGUF model loading

An opening report flags uncaught integer overflows and undefined behavior in the llama.cpp GGUF loader during model loading. No replies had appeared at the time of the report. Loaders that accept untrusted model files remain exposed until the overflows are closed.

Expert-parallel token dispatch for MoE models

A Hugging Face Transformers pull request adds expert-parallel token dispatch for mixture-of-experts models and makes it the default for Qwen3 MoE. CI failures were reported against the change. Distributed MoE inference gains an explicit dispatch path once the failures clear.

Maintainers debate AGENTS.md limits on AI agents

Two llama.cpp maintainers debated loosening strict AGENTS.md rules that block AI agents from writing pull requests or comments. The exchange is about project governance of automated contributions. Contributors who depend on agent tooling have a stake in whether the rules stay rigid.