trunk/7ca0e0bd5412c7d0a984f546a5e0a1a92f332313: [DTensor] skip decomposition for CIA ops (#174918)
This technical fix could dramatically speed up your PyTorch inference workloads...
Deep Dive
PyTorch developers have merged a critical fix (PR #174918) to address a performance regression in DTensor's CIA (Collective, Irregular, and Asynchronous) operations. The bug caused implicit redistributes during inference for ops like aten::linear, slowing down distributed computing workflows. The fix partially reverts changes from PR #171652, skipping decomposition flow when CIA decomposition is present to preserve the original, faster behavior that developers rely on for production systems.
Why It Matters
This fix prevents slowdowns in distributed PyTorch inference, directly impacting performance for large-scale AI deployments.