trunk/e6d6f0465ae435b4b73757553d3aa4504dd92d7d: [ROCm] forward fix #174087, take 4 (#175098)
A subtle PyTorch update broke major AI projects like vLLM on AMD hardware.
Deep Dive
A recent PyTorch commit (#175098) fixed a critical bug in the ROCm (AMD GPU) support that broke builds for the popular vLLM inference engine. The issue stemmed from a missing header inclusion for HIP stream functions masquerading as CUDA streams. This forward fix ensures compatibility for external projects relying on these low-level GPU abstractions, preventing build failures for developers using AMD hardware with frameworks like vLLM.
Why It Matters
This fix is crucial for maintaining stability in the AMD AI ecosystem, preventing major inference projects from breaking.