Unverified50% confidenceFactExact time
On an Apple Silicon Mac, after moving a PyTorch tensor on the MPS device to the CPU via .to("cpu"), the process's phys_footprint immediately drops to near zero, but the tensor data remains fully in memory.
1
Sources
50%
Confidence
Long-term
Relevance
10/13/2026
First Seen
Sources
Related Entities
Related Claims
UnverifiedIn the reproduction experiment (torch 2.14.0, macOS 26.5.2, M5 Pro 48 GB, 4 GiB tensor), right after moving to CPU both phys_footprint and the MPS driver count returned to near zero, and only after the CPU read through the copy page by page did the footprint return to 4.2 GiB, with the results of three runs completely identical.72% similarUnverifiedOn PC platforms, when VRAM is insufficient, part of the model can be offloaded to system memory, reducing speed but still allowing it to run; on Mac, when unified memory is insufficient, the model immediately reports an error and exits.69% similarUnverifiedThe llama.cpp project is optimized for Apple Silicon, making it possible to smoothly run 7B and even 13B quantized models on a Mac with 24GB unified memory.66% similarUnverifiedThe number of Ps defaults to the number of CPU cores on the machine, can be adjusted via the GOMAXPROCS environment variable, and when a queue is empty the scheduler will steal half the tasks from another P's queue.66% similarUnverifiedThe FP4-based model (svdq-fp4_r32) built on Nunchaku has a size of 6.6GB, fits entirely within 8GB of VRAM without CPU offloading, and reduces the per-step time to 3.4–3.8 seconds.65% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1015592API
curl https://kongchang.com/api/v1/knowledge/claims/1015592MCP
get_claim(id=1015592)