Worked on the pytorch/pytorch repository to enhance documentation for memory management metrics related to the pinned host allocator. Focused on improving observability and developer clarity, the update aligned documentation with the current implementation, detailing fields such as allocations, allocated_bytes, active_requests, and active_bytes. Clarified the semantics of counters and their relationship to memory events, reducing ambiguity in telemetry for both end users and internal tools. The work involved careful review of referenced pull requests and explicit notes on rounding and peak metrics. Utilized Python, CUDA, and technical writing skills to deliver these improvements without introducing any runtime changes.