from __future__ import annotations import importlib.util from contextlib import contextmanager from pathlib import Path import numpy as np import torch from psi.models.psi0 import ( ActionTransformerModel, Psi0Model, _validate_rtc_inference_contract, ) @contextmanager def _raises(error_type: type[Exception], message: str): try: yield except error_type as exc: assert message in str(exc), str(exc) else: raise AssertionError(f"Expected {error_type.__name__}: {message}") def _small_action_model(*, future_progress: bool = False) -> ActionTransformerModel: torch.manual_seed(7) model = ActionTransformerModel( attention_head_dim=4, num_attention_heads=2, caption_projection_dim=8, pooled_projection_dim=8, action_num_blocks=2, action_pred_horizon=6, action_dim=3, action_hidden_dim=8, action_nheads=2, n_conditions=0, odim=3, view_feature_dim=8, use_future_progress_embedding=future_progress, future_progress_dim=5, ) model.eval() return model def _forward_full_timeline( model: ActionTransformerModel, actions: torch.Tensor, *, future_progress: torch.Tensor | None = None, ) -> torch.Tensor: batch_size, horizon, _ = actions.shape return model( hidden_states=torch.empty(0), timestep=torch.full((batch_size, horizon), 500.0), joint_attention_kwargs={ "action_hidden_embeds": actions, "known_action_prefix": None, "views": torch.randn(batch_size, 1, 5, 8), "obs": torch.randn(batch_size, 1, 3), "traj2ds": None, "future_progress": future_progress, }, return_dict=True, ).action def test_v18_rtc_uses_one_full_timeline_action_forward() -> None: model = _small_action_model(future_progress=True) actions = torch.randn(2, 6, 3, requires_grad=True) output = _forward_full_timeline( model, actions, future_progress=torch.randn(2, 6, 5), ) assert output.shape == actions.shape # Stable initialization starts with an exact zero velocity field. torch.testing.assert_close(output, torch.zeros_like(output)) def test_stability_paths_are_zero_initialized_and_qk_normalized() -> None: model = _small_action_model(future_progress=True) for block in model.transformer_blocks: torch.testing.assert_close( block.norm1_act.linear.weight, torch.zeros_like(block.norm1_act.linear.weight), ) torch.testing.assert_close( block.norm1_act.linear.bias, torch.zeros_like(block.norm1_act.linear.bias), ) assert block.attn.norm_q is not None assert block.attn.norm_k is not None assert block.attn.norm_added_q is not None assert block.attn.norm_added_k is not None assert model.future_progress_mlp is not None torch.testing.assert_close( model.future_progress_mlp[-1].weight, torch.zeros_like(model.future_progress_mlp[-1].weight), ) torch.testing.assert_close( model.action_proj_out.linear.weight, torch.zeros_like(model.action_proj_out.linear.weight), ) def test_rtc_inference_delay_must_match_trained_overlap() -> None: assert _validate_rtc_inference_contract( action_horizon=30, action_exec_horizon=20, inference_delay=10, max_delay=11, ) == 10 with _raises(ValueError, "inference_delay=6, expected=10"): _validate_rtc_inference_contract( action_horizon=30, action_exec_horizon=20, inference_delay=6, max_delay=11, ) def test_rtc_model_entry_rejects_wrong_overlap_before_inference() -> None: model = Psi0Model.__new__(Psi0Model) torch.nn.Module.__init__(model) model.action_horizon = 30 model.action_dim = 3 model.action_exec_horizon = 20 model.rtc_max_delay = 11 model.device = torch.device("cpu") with _raises(ValueError, "inference_delay=6, expected=10"): model.predict_action_with_training_rtc_flow( observations=None, states=torch.zeros(1, 1, 3), instructions=["test"], num_inference_steps=1, traj2ds=None, prev_actions=torch.zeros(1, 30, 3), inference_delay=6, max_delay=11, ) def test_old_controller_shifts_exact_unexecuted_tail() -> None: deploy_path = ( Path(__file__).resolve().parents[1] / "deploy" / "psi_serve_rtc-trainingtimertc.py" ) spec = importlib.util.spec_from_file_location("psi_rtc_controller_test_module", deploy_path) assert spec is not None and spec.loader is not None module = importlib.util.module_from_spec(spec) spec.loader.exec_module(module) old_chunk = np.arange(30 * 3, dtype=np.float32).reshape(30, 3) shifted = module._shift_unexecuted_tail_to_prefix(old_chunk, 20) np.testing.assert_array_equal(shifted[:10], old_chunk[20:]) np.testing.assert_array_equal(shifted[10:], np.zeros((20, 3), dtype=np.float32)) if __name__ == "__main__": tests = [ value for name, value in sorted(globals().items()) if name.startswith("test_") and callable(value) ] for test in tests: test() print(f"psi0_rtc_test: {len(tests)} passed")