Commit History
-
Save Dita image preprocessing device with bundled runtime 2.2.0
Verified signatureSave CPU or policy-device FP32 resize/normalization for training and inference. Combine pixel checks on the model device. Keep cpu as the published default. Preserve all model weights, processors and action statistics. Bundle runtime 2.2.0, 372 passing tests, real-data loss/gradient/action comparisons in FP32/BF16/FP16, and archived runtime 2.1.0 metadata.0509d0a
-
Update Dita training configuration and bundled runtime to 2.1.0
Verified signatureEnable full-backbone fine-tuning and inference checkpoint bypass; add native LoRA support. Align optimizer groups and the warmup/cosine schedule with the upstream LIBERO recipe, including LIBERO-Long's higher learning rate. Preserve all model weights and processors. Bundle the tested 2.1.0 wheel, source hashes, validation reports, and archived 2.0.1 metadata.40690c3
-
Use policy.dtype as the sole precision setting (Dita 2.0.1)
Verified signatureRemove use_amp from model config.json and publish the validated 2.0.1 runtime. The inherited LeRobot flag is derived internally from dtype. Model weights, processor files, statistics and gripper modes are unchanged.5a61b42 -
Add binary/continuous gripper modes with DITA runtime 2.0.0
Verified signatureUpdate policy/processor configs, runtime wheel, documentation and real-data validation. Preserve model weights and LIBERO statistic tensors.3828ef7 -
Add Dita runtime 1.0.5 and native policy.dtype support
Verified signaturePublish the validated runtime wheel, explicit legacy FP16 compute default, model-card installation commands and runtime verification reports. Model weights and processor files are unchanged.e222ad7
-
Synchronize Dita libero_10_no_noops LeRobot checkpoint
Verified signature68a9edd -
Convert legacy Dita long checkpoint
Verified signature4fa5f72