PaddlePaddle fixes NVSHMEM dependency gate for FlashMask on Hopper GPUs

{
  "event_id": "event_aede816a39a6d2a2",
  "topic_id": "gpu_supply_and_compute_market",
  "event_type": "Pull Request",
  "event_time": "2026-06-05T00:53:30Z",
  "title": "PaddlePaddle fixes NVSHMEM dependency gate for FlashMask on Hopper GPUs",
  "summary": "A previous commit accidentally disabled the NVSHMEM communication library for all non-XPU builds, causing a runtime crash in FlashMask on Hopper GPUs. This PR corrects the CMake logic to only enable NVSHMEM when FlashAttention is active on the target architecture.",
  "contribution": "Corrects a build configuration regression that blocked NVSHMEM for FlashMask on Hopper GPUs, restoring correct build dependencies for distributed training.",
  "impact": "Developers using PaddlePaddle for distributed training with FlashMask on Hopper GPUs can now build and run without hitting a missing library crash. This fixes a specific regression that broke CI coverage tests. It's important to watch if this more precise dependency gate holds for other FlashAttention features or future GPU architectures.",
  "maturity": "Code",
  "confidence": 0,
  "importance_score": 0.65,
  "risk_flags": [
    "Build Configuration Dependency",
    "Regression Fix"
  ],
  "evidence_count": 1
}