PyTorch 2.14.0 ships, security docs clarified

What shipped: PyTorch 2.14.0 released and went live in the release matrix [1][2]. Alongside, two security documentation changes clarify that checkpoints share the distributed-features trust model and update security guidelines; review these if you handle untrusted checkpoints [3][4]. Also landed: MPS complex std/var now returns real results [5], and torch::nn::Sequential::clone() preserves submodule names [6]. Inductor fixed stale captured-tensor mutations in while_loop, and executorch reports NotFound when a CUDA delegate's weights blob is missing, plus fixed dropped edge-manager passes [7][8]. Torchtitan added a B200 Kimi K3 CI lane, fixed RL CIFlow tags [9], and overhauled MxFP8 support [9].

Why it matters: The import reload rejection means code that catches an initial failed import and retries will break; PyTorch 2.14.0's CUDA graph freeing now relies on refcounts, not the cyclic GC, which could affect memory management in long-running processes [10][11]. Torchtitan now trains MoE routers in fp32 with bf16x9 on Blackwell, a shift if you rely on prior numerics [9]. The security doc clarifications make it explicit that checkpoints are not a safe boundary; treat them with the same suspicion as distributed features [3].

Under the hood: The MPS fix corrects the dtype of std/var results for complex tensors, aligning with CPU/CUDA behavior [5]. CUDA graph annotation leaks on failed capture were fixed, and the clear_kernel_annotations workflow retired [14]. Executorch also runs a coalesced TensorRT + CUDA program in the reference runner gate [8]. Torchtitan counts both attention matmuls in FLOPs estimates and uses the actual step window for end-to-end timing [12][13].

Also: Test infrastructure added an AI_NOT_RELATED bucket for advisor-cleared failures [15], and ROCm removed outdated version conditions [16].