PyTorch: [inductor] Extend partitioned scatter to additional ops (#194078)
The partitioned scatter pass only matched aten.index_put / aten.index_put_ , so it missed ops that emit the exact same contended atomic_add store.
Executive Summary
The partitioned scatter pass only matched aten.index_put / aten.index_put_ , so it missed ops that emit the exact same contended atomic_add store. This extends the pass to cover them and tightens the partition cap so the rewrite cannot cost more memory traffic than the scatter it replaces. Adds support for: _unsafe_masked_index_put_accumulate lowering aten.scatter_add , aten.scatter_reduce.two , aten.scatter.reduce lowering Traffic-aware partition cap.** The old cap was P <= 4 * writes_per_slot , justified as diminishing returns.
Hyperscale Open Source & Cloud Infrastructure Review
From an infrastructure engineering, cloud-native scale, and hyperscale governance standpoint: - **Hyperscale Provenance:** Open-sourcing internal frameworks subjects proprietary systems to rigorous public analysis. - **De-facto Standard Cohesion:** Publishing enterprise tooling establishes interoperable specifications across distributed execution. - **Vendor Decoupling:** Platform engineers can inspect underlying telemetry and memory layouts without black-box vendor lock-in.
Impact on the Open Ecosystem
Bridges enterprise engineering scale with independent, reproducible open-source software stacks.