June 2026 monthly summary for the pytorch/FBGEMM project focusing on hardening host-mapped tensor support in UVM transfers. Delivered a new CUDAHostMappedIndirectContext to safely manage host-mapped tensors during uvm_to_cpu/uvm_to_device operations, preventing null pointer dereferences and improving memory safety. Implemented the corresponding deleter branches and CPU/Device view construction in memory_utils.cu to align with the established managed-tensor pattern. Added end-to-end tests to validate behavior under large tensor allocations, increasing reliability for production workloads. Resolved a production crash path and reduced runtime variance in critical checkpointing/preranker workflows, eliminating tensor-zero compact_stalls across multiple nodes. Demonstrated strong CUDA memory-management, C++/CUDA development, and test-driven development skills, with effective cross-repo collaboration.