AI toolchain bugs and backend kernels

PyTorch Inductor drew a security-tinged bug report on silent NaN output from zero-size interpolate, while llama.cpp and Hugging Face Transformers saw hardware kernel and training-control proposals. Smaller feature requests for FP8, matmul folding, and multimodal UI rounded out the day.

Inductor NaN on zero-size interpolate

A pytorch/pytorch bug report states that Inductor produces NaN via an out-of-bounds read when F.interpolate meets a zero-size spatial dimension, whereas eager mode raises. Output size 0 is accepted silently. The mismatch matters for torch.compile users who hit edge-case shapes and risk silent data corruption.

Intel Xe flash attention for llama.cpp Vulkan

A ggml-org/llama.cpp pull request adds Intel Xe-specific flash attention kernels to the Vulkan backend for Xe-LPG Plus, Xe2, and Xe3, accompanied by long-context benchmarks. Local LLM runners on recent Intel GPUs gain a concrete performance path for attention-heavy workloads.

Pause mechanism for continuous batching

A huggingface/transformers pull request introduces a pause mechanism for continuous batching. The author reports it supports interleaved generation and training under data and tensor parallelism with minimal overhead. Distributed training users who mix generation steps can keep pipelines responsive without large cost.

Bool mask and hidden state for Gemma4

A huggingface/transformers pull request adds bool mask support for logits_to_keep and exposes last_hidden_state in Gemma4ForConditionalGeneration. Two developers frame the changes as memory-saving measures. Gemma users who need tighter control over generation footprints have a direct path to lower peak usage.

Inline images in llama.cpp WebUI

A ggml-org/llama.cpp feature request asks for inline image rendering inside chat messages, including MCP tool outputs, in the WebUI. The goal is better multimodal tool display. Multimodal local-inference users watching WebUI polish will track whether the request lands.

FP8 scalar scatter CUDA kernel

A pytorch/pytorch issue requests a missing FP8 scalar scatter CUDA kernel. Completing the kernel would close a gap in low-precision support. FP8 workloads on NVIDIA hardware become more complete once the primitive is present.

Matmul folding for non-contiguous views

A pytorch/pytorch issue asks Inductor to support matmul folding for zero-copy non-contiguous views under torch.compile. The change would let more view patterns stay optimized. Compile users with non-contiguous layouts stand to keep performance that currently falls back.

Vulkan tensor parallelism tracking

A ggml-org/llama.cpp issue opens tracking for Vulkan tensor parallelism support. It frames multi-device scaling on the Vulkan backend. Contributors focused on multi-GPU local inference have a place to follow progress.