openpi
openpi copied to clipboard
Physical-Intelligence
I use the command: uv run scripts/serve_policy.py policy:checkpoint --policy.config=pi05_libero --policy.dir=gs://openpi-assets/checkpoints/pi0_base but it throws an error: INFO:root:Loading model... INFO:2025-10-21 05:46:22,801:jax._src.xla_bridge:925: Unable to initialize backend 'rocm': module 'jaxlib.xla_extension' has no attribute 'GpuAllocatorConfig'...
XLA_PYTHON_CLIENT_MEM_FRACTION=0.9 uv run scripts/train.py mytest --exp-name=my_experiment --overwrite This command seems to perform full fine-tuning. How can I modify it to perform LoRA fine-tuning instead?
Seems like it should be config.action_dim instead of 32 for line 100, 101 and 107 ([link](https://github.com/Physical-Intelligence/openpi/blob/95aadc6b16d170e4b13ab2e0ac64fbf2d1bb8e31/src/openpi/models_pytorch/pi0_pytorch.py#L10)).
I noticed that the open-source pi0.5 was trained using the knowledge insulation method, rather than the initial approach described in the original pi0.5 paper. Is it possible to release a...
## What's this This enables to develop this repository on macOS.
When I fine-tuning the pi05, it always stopped when save the model. Here for debug I set save interval equals 10, an it still stop the process without any error...
I have set up pi0 model in a remote server and aloha sim in my local device. Basically, I have a script to get the observation from the simulator each...
when i tried to install `GIT_LFS_SKIP_SMUDGE=1 uv pip install -e .` ``` Computing main repo mapping: Computing main repo mapping: Loading: Loading: 0 packages loaded Loading: 0 packages loaded WARNING:...
--------------------------------------------------------------------------- ValueError Traceback (most recent call last) Cell In[2], [line 1](vscode-notebook-cell:?execution_count=2&line=1) ----> [1](vscode-notebook-cell:?execution_count=2&line=1) policy = policy_config.create_trained_policy(config, checkpoint_dir) File /data/wuyuzhou-20250917/openpi/src/openpi/policies/policy_config.py:54, in create_trained_policy(train_config, checkpoint_dir, repack_transforms, sample_kwargs, default_prompt, norm_stats, pytorch_device) 52 logging.info("Loading model...")...
