Security rejections and GPU path work in llama.cpp and PyTorch

Security issues led the day's AI tooling traffic, with a llama.cpp container proposal rejected over host filesystem exposure and a PyTorch Inductor sizing flaw flagged for silent corruption. Parallel threads advanced Intel GPU kernels and graph replay in llama.cpp, alongside smaller model-support and ABI requests in Ollama, Transformers, and PyTorch.

llama.cpp rejects Apptainer support over filesystem bypass

A pull request to add Apptainer container support to the llama.cpp server drew a single rejecting comment. The reviewer stated the change bypasses validation and permits arbitrary host filesystem access through the tools endpoint. Operators who rely on container boundaries for local model serving have a concrete reason to avoid the proposed path.

PyTorch Inductor split-scan workspace overrun risks silent corruption

An opening bug report in pytorch/pytorch describes how Inductor split-scan workspace sizing can be overrun by an undersized R0_BLOCK. The mismatch can produce silent memory corruption. Users of Inductor-generated kernels on the affected scan path should watch the report for corrective sizing logic.

Intel Xe Vulkan flash-attention kernels proposed for llama.cpp

A pull request adds flash-attention optimization kernels for Intel Xe under the llama.cpp Vulkan backend, covering Xe-LPG Plus, Xe2, and Xe3. Accompanying benchmark data shows gains on long-context workloads. The change targets better inference throughput on recent Intel integrated GPUs.

SYCL graph record and replay added to llama.cpp

A pull request introduces SYCL graph record and replay support in llama.cpp. The thread contains early Intel GPU benchmarks and tuning questions from five participants. The feature is intended to cut launch overhead for repeated kernel sequences on the SYCL backend.

Bool mask and last_hidden_state exposure for Gemma4 in Transformers

A pull request in huggingface/transformers adds bool mask support for logits_to_keep and exposes last_hidden_state on Gemma4ForConditionalGeneration. Limited discussion so far focuses on the API surface. Callers that need selective logit retention or direct final-state access would benefit if the change merges.

Native spark2_5 architecture support requested in Ollama

A user asked for native Ollama support of the spark2_5 architecture used by Spark-X2.5 4B and 1.7B models. Those weights currently require a custom llama.cpp fork. Built-in handling would remove the fork dependency for deployers of the models.

Compile-time failure for oversized TORCH_TARGET_VERSION

A PyTorch issue requests that a TORCH_TARGET_VERSION newer than the built Torch version produce a compile-time error under the Stable ABI. The single-message proposal aims at clearer diagnostics for extension authors. Catching the mismatch at build time would reduce runtime surprises.

Config flag to decouple last_hidden_state in Transformers

A pull request proposes a model config option to untie hidden_states[-1] from last_hidden_state in huggingface/transformers. The single-comment change would give callers explicit control over output identity. Code that currently assumes the two tensors are the same object would need to respect the new flag.