from __future__ import annotations from pathlib import Path import numpy as np import torch from psi.data.seed.full_duplex_audio_command_dataset import ( AudioSource, FullDuplexAudioCommandDataset, HiddenSlice, ) from psi.models.full_duplex_psi0 import StreamingCausalMemory def _minimal_source_dataset(monkeypatch) -> FullDuplexAudioCommandDataset: dataset = FullDuplexAudioCommandDataset.__new__(FullDuplexAudioCommandDataset) dataset.records = [ { "row_id": 3, "filename": "test_motion", "instruction": "test", "audio_dir": "seed_audio/000003", "motion_dir": Path("/unused/test_motion"), "declared_frames": 20, } ] dataset.audio_catalog = { "seed_audio/000003": ( AudioSource( "seed_audio/000003/voice", (HiddenSlice(0, Path("/unused/0"), 0, 1),), ), ) } dataset.white_noise_hidden = (torch.full((2, 8), 9.0),) dataset.fixed_voice_per_record = True dataset.action_chunk_size = 30 dataset.action_exec_horizon = 20 dataset.prefix_frames = 10 dataset.max_motion_steps = 0 dataset.initial_idle_max_steps = 1 dataset.stationary_blend_frames = 20 motion = np.zeros((20, 36), dtype=np.float32) motion[:, 32] = 1.0 monkeypatch.setattr(dataset, "_load_motion", lambda path: motion) monkeypatch.setattr(dataset, "_load_slice", lambda item: torch.ones(2, 8)) monkeypatch.setattr( dataset, "_sample_stationary_anchor", lambda task_start_pose: np.asarray(task_start_pose, dtype=np.float32), ) monkeypatch.setattr( dataset, "_sample_stationary_sequence", lambda phase, frame_count, anchor_pose: ( np.repeat( np.asarray(anchor_pose, dtype=np.float32)[None, :], frame_count, axis=0, ), f"test-{phase}", ), ) monkeypatch.setattr( dataset, "_build_task_sequence", lambda task_motion, stationary_anchor: ( np.asarray(task_motion, dtype=np.float32), 1, ), ) monkeypatch.setattr("random.randint", lambda low, high: 1) return dataset def test_idle_listen_and_motion_keep_qwen_cadence_with_dense_actions(monkeypatch): source = _minimal_source_dataset(monkeypatch)[0] phases = [step["phase"] for step in source["steps"]] assert phases.count("idle") == 20 assert phases.count("listen") == 20 assert phases.count("motion") == 9 assert [ index for index, step in enumerate(source["steps"]) if step["update_semantic_cache"] ] == [0, 20, 40] assert all( step["action_substep"] == 0 for step in source["steps"] if step["update_semantic_cache"] ) assert all( not step["update_semantic_cache"] for step in source["steps"] if step["action_substep"] != 0 ) def test_continuous_cache_caps_each_source_without_padding_shrinkage(): memory = StreamingCausalMemory( input_dim=8, hidden_dim=16, num_layers=1, num_heads=4, dropout=0.0, max_cache_tokens=3, ).eval() cache = memory.begin_source(["long-source", "finished-source"]) for _ in range(3): _, cache, _ = memory( torch.randn(2, 8), cache, source_ids=["long-source", "finished-source"], update_cache=[True, True], ) assert cache.sequence_lengths == (3, 3) # False denotes only a padded row. Continuing the other source must retain # the finished source's newest three real observations under the shared cap. for _ in range(4): _, cache, _ = memory( torch.randn(2, 8), cache, source_ids=["long-source", "finished-source"], update_cache=[True, False], ) assert cache.sequence_lengths == (3, 3) assert cache.valid_mask is not None assert cache.valid_mask.sum(dim=-1).tolist() == [3, 3] def test_all_padding_rows_leave_cache_objects_unchanged(): memory = StreamingCausalMemory( input_dim=8, hidden_dim=16, num_layers=2, num_heads=4, dropout=0.0, max_cache_tokens=4, ).eval() cache = memory.begin_source(["source-a", "source-b"]) _, cache, _ = memory( torch.randn(2, 8), cache, source_ids=["source-a", "source-b"], update_cache=[True, True], ) before_keys = cache.keys before_values = cache.values _, padded_cache, _ = memory( torch.randn(2, 8), cache, source_ids=["source-a", "source-b"], update_cache=[False, False], ) assert padded_cache.sequence_lengths == (1, 1) assert all(before is after for before, after in zip(before_keys, padded_cache.keys)) assert all(before is after for before, after in zip(before_values, padded_cache.values))