PyTorch memory bugs and backend proposals

PyTorch saw two memory-safety reports in Inductor and CUDA kernels, while Transformers and llama.cpp advanced parallelism and backend work. A heated review of AI-generated comments and several RFCs filled out the rest of the day.

Inductor CPU 2D reduction tail overflow

PyTorch Inductor CPU codegen for 2D-tiled reductions stores a full-width vector in the tail block. The defect can trigger a heap overflow that aborts with SIGABRT or produce silent wrong results. Anyone running CPU reductions through the compiler needs correct tail handling to avoid corruption.

CUDA DeviceScanKernel out-of-bounds read

A bug report describes an illegal memory read inside PyTorch CUDA DeviceScanKernel when performing cumsum on tensors larger than 2^30 elements. The issue lives in the pytorch/pytorch tracker and surfaces as device-side out-of-bounds access. Large cumulative-sum workloads on GPU are directly exposed.

FSDP2 plus expert parallelism via 2-D mesh

An umbrella pull request in huggingface/transformers opens coordinated work to enable FSDP2 together with expert parallelism through a 2-D device mesh of fsdp and tp. The single-message proposal starts the design track for that combination. Distributed training of mixture-of-experts models is the immediate audience.

GLM-5.3-Flash PR criticized for AI comments

A pull request adding GLM-5.3-Flash (glm5next) to ggml-org/llama.cpp drew a heated complaint that its AI-generated comments violate project rules as slop. Two participants exchanged views on the submission. Model-integration reviewers and contributors tracking comment standards are affected.

Dynamo graph breaks in checkpointed functions

A design note in pytorch/pytorch outlines support for graph breaks inside torch.checkpoint by introducing per-region higher-order operators. The single message sketches how Dynamo would handle the breaks. Users combining activation checkpointing with torch.compile have a stake in the outcome.

Radix-select top_k for SYCL

An RFC in ggml-org/llama.cpp proposes a radix-select implementation of top_k for the SYCL backend and plans to move the code into separate files after review feedback. The change targets more efficient top-k on SYCL devices. SYCL backend maintainers and users should watch the direction.

OpenMP barrier slowdown on mingw-w64

An issue in ggml-org/llama.cpp reports roughly 40 percent decode cost when GGML_OPENMP is enabled under mingw-w64 libgomp, because every graph barrier becomes a non-spinning kernel semaphore. The reporter suggests documentation updates or a default change. Windows builds that rely on OpenMP for small-op-heavy graphs are hit.

SYCL graph record and replay

A pull request adds SYCL graph record and replay support to ggml-org/llama.cpp. Early discussion seeks reproduction help and tuning advice. The feature is intended to cut launch overhead for repeated SYCL execution graphs.