Spaces:

optimum

/

llm-perf-leaderboard

like 336

Running

There are multiple ways to use CUDA memory, pytorch for example will allocate some memory for tensors, but will also reserve some for its computation, so reserved = (allocated + cached). It's important to look at both because the performance we observe depends on that reserved memory. that's also why sometimes you can load a model but OOM when you run it.
Finally the used~= (resrved + non-releasable) is technically what you'll observe on nvidia-smi, the most external view of memory usage.
More in https://pytorch.org/docs/stable/generated/torch.cuda.memory_stats.html