Nikita Shulga 6db1b9dd21 [MPS] Chunk fillBuffer into 4Gb slices (#164108) ...

To avoid regression on MacOS 26, which one could observe by running the following script
```swift
import Metal

let bufferSize = 1<<32 + 4

guard let device = MTLCreateSystemDefaultDevice() else { fatalError("No Metal device found") }
guard let buffer = device.makeBuffer(length: bufferSize, options: .storageModeShared) else { fatalError("Failed to create buffer") }

guard let cmdQueue = device.makeCommandQueue() else { fatalError("Failed to create command queue") }
guard let cmdBuffer = cmdQueue.makeCommandBuffer() else { fatalError("Failed to create command buffer") }
guard let blitEncoder = cmdBuffer.makeBlitCommandEncoder() else { fatalError("Failed to create blit encoder") }

blitEncoder.fill(buffer: buffer, range: 0..<bufferSize, value: 0x42)
blitEncoder.endEncoding()

cmdBuffer.commit()
cmdBuffer.waitUntilCompleted()

let tailOffs = 8
let hostPtr = buffer.contents().bindMemory(to: UInt8.self, capacity: bufferSize)
let tail = Array(UnsafeBufferPointer(start: hostPtr + (bufferSize - tailOffs), count: tailOffs))

for (idx, val) in tail.enumerated() {
    print("Offs 0x\(String(bufferSize - tailOffs + idx, radix: 16)): 0x\(String(val, radix: 16))")
}
```

Test plan: run `test_indexing.py` on MacOS-26

Fixes https://github.com/pytorch/pytorch/issues/161265
Pull Request resolved: https://github.com/pytorch/pytorch/pull/164108
Approved by: https://github.com/Skylion007

2025-09-29 20:19:29 +00:00

..

conda

PyTorch -> C++17 (#98209) (#100557)

2023-05-19 00:49:08 +00:00

src

[MPS] Chunk fillBuffer into 4Gb slices (#164108)

2025-09-29 20:19:29 +00:00

tools

[BE][CI] Get rid of duplicated code (#131406)

2024-07-23 04:01:13 +00:00

CMakeLists.txt

Revert "Use official CUDAToolkit module in CMake (#154595)"

2025-06-23 21:15:31 +00:00