pub fn set_param_sharing_scope(label: &str)Expand description
Declare which compiles belong to the same model, for backends that can share parameter storage between executables.
A model that compiles several graphs (a prefill graph plus decode buckets)
otherwise re-uploads every weight per graph. Sharing that storage means
skipping the re-upload, which is only sound when the same parameter name and
size really is the same tensor — and nothing in a graph establishes that,
since parameter content arrives later via set_param. So the caller
declares it: pass a stable label per model (a weights path works well).
No-op unless a backend that supports sharing is compiled in. Today that is
CUDA, additionally gated behind RLX_CUDA_SHARED_PARAMS=1.