Small fixes
at the runtime boundary.

LLVM / MLIR

LLVM

4 merged PRs

MLIR is the compiler infrastructure a lot of ML tooling is built on. These fixes are all the same shape: a transform assumed an input it can't actually count on — a rank match, a return statement, an attribute — and crashed instead of just handling it.

• [mlir][SCF] Fix use-after-free in coalesceLoops when inner loop yields its induction var

  Merging two loops into one could leave a dangling reference if the inner loop reused its own counter afterward — fixed the memory bug, not just the crash it caused.

• [mlir][Bufferization] Don't assert in foldMemRefCasts on a func body without func.return

  Compiling a function with no return statement crashed this optimization pass outright, instead of just skipping it.

• [mlir][Ptr] Don't assert when the default memory space isn't a MemorySpaceAttrInterface

  A pointer-handling pass assumed every memory space carried extra metadata it doesn't always have — it crashed instead of falling back gracefully.

• [mlir][SparseTensor] Fix crash demapping alloc_tensor with a copy operand

  Allocating a sparse tensor with a copy attached hit a code path the sparse-layout pass never expected, and crashed the compiler.

Megatron-LM

NVIDIA

1 merged PR

Large-scale transformer training. One config-validation check had drifted out of sync with the code it was supposed to mirror — fixed so it actually catches what it claims to.

• Sync yaml VPP pipeline-parallel-size check with arguments.py

  A YAML config validator and the real argument-parsing logic had quietly drifted apart, so the validator could wrongly pass or fail a config — synced the two back up.

arapuca

LeGambiArt

2 merged PRs

A process sandbox for Linux, macOS, and Windows. Both merged fixes closed fail-open gaps — cases where the sandbox looked like it was isolating a program but quietly wasn't.

• fix(windows): always engage AppContainer; make LPAC opt-in

  On Windows the sandbox could skip its strongest isolation mode unless a weaker one was explicitly set — flipped the default so the strong isolation is always on.

• fix(sandbox): don't grant read access to host /tmp by default

  Sandboxed programs could silently read the real host machine's /tmp directory by default, even though the whole point of the sandbox is that they can't.

OpenRLHF

OpenRLHF

1 merged PR

A Ray + vLLM RLHF training framework. Checkpoint saving broke for any LoRA-adapted model with tied embeddings — the fix was a one-line matching bug with an outsized blast radius.

• fix(deepspeed): match tied lm_head key by suffix under PEFT wrapping in save_model

  Saving a checkpoint under DeepSpeed ZeRO-3 with a LoRA adapter on a tied-embedding model (e.g. Qwen2.5) crashed with a 'mismatch keys' assertion — PEFT renames the tied weight and the special-case check only matched the bare name. Matched by suffix instead so it finds the key regardless of wrapping.

Unsloth

unslothai

1 merged PR

Fast LLM fine-tuning, including the Unsloth Studio desktop app. A wildcard server bind handed the wrong IP to the sharing banner — a privacy bug, not just a display bug.

• fix(studio): stop handing a wildcard bind's public IP to LAN peers

  Binding Studio's server to 0.0.0.0 made it show your public internet IP as the address to share with 'another device on your network' — a LAN device usually can't even route to that, and it needlessly exposed the public IP. Resolved the real LAN address instead.