Repository navigation
This release improves training stability, normalization configuration, and mjlab integration, with fixes for hook initialization, CLI argument inheritance, and package distribution.
Added
NormFactoryfor configurable RMSNorm and LayerNorm in attention and transformer layers, includingepsandelementwise_affineoptions forqk_normandblock_norm.- Custom activation imports through module-qualified names in
resolve_activation_fn. dtype,copy, andrequires_gradoptions inAgent.to_tensor.- Public export of
configure_distributed_cache_dirsfromcusrl.utils; the helper now safely skips configuration whenRANKis unset.
Breaking / Migration Notes
HookComposite.compile()now leaves the composite objective uncompiled by default. Passcompiled_objective=Trueto enable its compilation.- Activation strings such as
"nn.ReLU"and"torch.ReLU"should be replaced with"ReLU"or"torch.nn.ReLU".
Fixed
- Improved numerical stability in
AdvantageNormalizationby using population variance, avoiding undefined variance for single-sample inputs. - Improved grouped variance computation in
RunningMeanStdand added validation for positiveepsilon. ObservationNormalizationnow defers synchronization only when distributed execution is enabled.- Initially inactive hooks now receive
pre_init,init, andpost_init, ensuring they are ready for later activation. Agent.to_tensornow copies NumPy input by default, preventing unintended storage sharing.- Fixed inheritance of inline Tyro arguments such as
--env-args=...and--env-kwargs=.... EnvironmentFactorySpec.make_environment()no longer mutates the stored environment kwargs.MjlabEnvAdapternow respects the environment'sauto_resetsetting, reports its timestep, and saves/restorescommon_step_counter.- Added positive
num_batchesandbatch_sizevalidation toRandomSampler,TemporalRandomSampler, andAutoRandomSampler. - Corrected setuptools package discovery to include all
cusrlsubpackages.