MACE0.3.16 two native CPU format conversions with cuequivariance0.7.0,restored8ownSi frames E/F/stress+weights114checks. OriginalactualGPUcandidateFALSEpreserved:exactopswheel CUBINmin75/PTX100/120 lacksP10060/V10070;noCUDAcueqforward/accelerationclaim. CPUtoolconversion usesfrontendnaiveTorch,notGPUfallbackcredit. --return_model actionstore_false causes SAVEinthisversion. Dependenciesprivate6pins;no NVIDIAvendorwheels/binaries/source redistributed. Ownmodelweights/data only. bashreadplotarchived;optionalsetup.sh thenCPU sbatchsubmit.
