Merge remote-tracking branch 'origin/main' into worktree-lingbot-va-port

# Conflicts:
#	docs/source/_toctree.yml
#	src/lerobot/policies/factory.py
#	uv.lock

tests/policies/fastwam/test_fastwam_policy.py

@@ -0,0 +1,391 @@
+#!/usr/bin/env python
+
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import json
+
+import pytest
+import torch
+from safetensors import safe_open
+from torch import nn
+
+pytest.importorskip("transformers", reason="fastwam requires the `fastwam` extra (transformers)")
+pytest.importorskip("diffusers", reason="fastwam requires the `fastwam` extra (diffusers)")
+
+from lerobot.configs import FeatureType, PolicyFeature, PreTrainedConfig
+from lerobot.policies import FastWAMConfig, get_policy_class, make_policy_config, make_pre_post_processors
+from lerobot.policies.fastwam.modeling_fastwam import FastWAMPolicy
+from lerobot.policies.fastwam.processor_fastwam import FastWAMActionToggleProcessorStep
+from lerobot.utils.constants import ACTION, OBS_STATE
+
+
+class FakeFastWAMCore(nn.Module):
+    def __init__(self):
+        super().__init__()
+        self.dit = nn.Linear(2, 2)
+
+    def training_loss(self, sample):
+        assert sample["video"].ndim == 5
+        assert sample["context"].ndim == 3
+        return sample[ACTION].sum() * 0.0 + torch.tensor(1.0), {"loss_action": 1.0}
+
+    def infer_action(self, **kwargs):
+        return {"action": torch.ones(1, kwargs["action_horizon"], 3)}
+
+
+def test_fastwam_is_registered_and_publicly_exported():
+    cfg = make_policy_config(
+        "fastwam",
+        action_dim=3,
+        proprio_dim=2,
+        action_horizon=4,
+        n_action_steps=2,
+        num_video_frames=5,
+        action_video_freq_ratio=1,
+        base_model_id=None,
+    )
+
+    assert isinstance(cfg, FastWAMConfig)
+    assert cfg.type == "fastwam"
+    assert get_policy_class("fastwam") is FastWAMPolicy
+
+
+def test_config_validates_features_model_ids_and_saved_auto_route(tmp_path):
+    cfg = FastWAMConfig()
+    cfg.save_pretrained(tmp_path)
+    saved = json.loads((tmp_path / "config.json").read_text())
+
+    assert saved["pretrained_path"] is None
+    assert cfg.image_features["observation.images.image"].type == FeatureType.VISUAL
+    assert cfg.action_feature.shape == (7,)
+    assert cfg.robot_state_feature.shape == (8,)
+    with pytest.raises(ValueError, match="image feature"):
+        FastWAMConfig(input_features={OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(8,))})
+    assert FastWAMConfig(tokenizer_model_id="somebody/other-tokenizer").tokenizer_model_id == (
+        "somebody/other-tokenizer"
+    )
+
+
+def test_preprocessor_passes_images_through_and_postprocessor_toggles_actions(tmp_path):
+    cfg = FastWAMConfig(
+        action_dim=3,
+        proprio_dim=2,
+        action_horizon=4,
+        n_action_steps=2,
+        num_video_frames=5,
+        action_video_freq_ratio=1,
+        image_size=(2, 2),
+        device="cpu",
+        toggle_action_dimensions=[-1],
+        input_features={
+            "observation.images.image": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 2, 2)),
+            OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(2,)),
+        },
+        output_features={ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(3,))},
+        base_model_id=None,
+    )
+    dataset_stats = {
+        "observation.images.image": {
+            "mean": torch.full((3, 1, 1), 0.2),
+            "std": torch.full((3, 1, 1), 0.1),
+        },
+        OBS_STATE: {
+            "mean": torch.tensor([1.0, 3.0]),
+            "std": torch.tensor([2.0, 4.0]),
+        },
+        ACTION: {
+            "mean": torch.zeros(3),
+            "std": torch.ones(3),
+        },
+    }
+
+    preprocessor, postprocessor = make_pre_post_processors(cfg, dataset_stats=dataset_stats)
+    processed = preprocessor(
+        {
+            "observation.images.image": torch.tensor(
+                [
+                    [[0.0, 0.5], [1.0, 0.5]],
+                    [[0.0, 0.5], [1.0, 0.5]],
+                    [[0.0, 0.5], [1.0, 0.5]],
+                ]
+            ),
+            OBS_STATE: torch.tensor([3.0, 7.0]),
+        }
+    )
+    preprocessor.save_pretrained(tmp_path, config_filename="policy_preprocessor.json")
+    postprocessor.save_pretrained(tmp_path, config_filename="policy_postprocessor.json")
+    _, loaded_postprocessor = make_pre_post_processors(cfg, pretrained_path=str(tmp_path))
+
+    # VISUAL normalization is IDENTITY
+    expected_image = torch.tensor(
+        [[[[0.0, 0.5], [1.0, 0.5]], [[0.0, 0.5], [1.0, 0.5]], [[0.0, 0.5], [1.0, 0.5]]]]
+    )
+    assert preprocessor.name == "policy_preprocessor"
+    assert postprocessor.name == "policy_postprocessor"
+    assert torch.allclose(processed["observation.images.image"], expected_image)
+    assert torch.allclose(processed[OBS_STATE], torch.tensor([[1.0, 1.0]]))
+    assert torch.equal(dataset_stats["observation.images.image"]["mean"], torch.full((3, 1, 1), 0.2))
+    assert any(isinstance(step, FastWAMActionToggleProcessorStep) for step in loaded_postprocessor.steps)
+    assert torch.equal(
+        loaded_postprocessor(torch.tensor([[0.25, 0.5, 1.0]])), torch.tensor([[0.25, 0.5, -1.0]])
+    )
+
+
+def test_policy_forward_and_predict_action_adapt_lerobot_batches(monkeypatch):
+    captured = []
+
+    class CapturingCore(FakeFastWAMCore):
+        def infer_action(self, **kwargs):
+            captured.append(
+                {
+                    "image_shape": tuple(kwargs["input_image"].shape),
+                    "proprio_shape": tuple(kwargs["proprio"].shape),
+                    "prompt": kwargs["prompt"],
+                }
+            )
+            return {"action": torch.full((1, kwargs["action_horizon"], 3), float(len(captured)))}
+
+    monkeypatch.setattr(FastWAMPolicy, "_build_core_model", lambda self, config: CapturingCore())
+    cfg = FastWAMConfig(
+        action_dim=3,
+        proprio_dim=2,
+        action_horizon=4,
+        n_action_steps=2,
+        num_video_frames=5,
+        action_video_freq_ratio=1,
+        image_size=(16, 16),
+        input_features={
+            "observation.images.image": PolicyFeature(type=FeatureType.VISUAL, shape=(3, 16, 16)),
+            OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(2,)),
+        },
+        output_features={ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(3,))},
+        base_model_id=None,
+    )
+    policy = FastWAMPolicy(cfg)
+
+    loss, metrics = policy.forward(
+        {
+            "observation.images.image": torch.zeros(1, 3, 16, 16),
+            OBS_STATE: torch.zeros(1, 2),
+            ACTION: torch.zeros(1, 4, 3),
+            "context": torch.zeros(1, 5, 4096),
+            "context_mask": torch.ones(1, 5, dtype=torch.bool),
+        }
+    )
+    action = policy.predict_action_chunk(
+        {
+            "observation.images.image": torch.stack(
+                [
+                    torch.zeros(3, 16, 16),
+                    torch.ones(3, 16, 16),
+                ]
+            ),
+            OBS_STATE: torch.tensor([[0.0, 1.0], [2.0, 3.0]]),
+            "task": ["task 0", "task 1"],
+        }
+    )
+
+    assert loss.item() == 1.0
+    assert metrics["loss_action"] == 1.0
+    assert action.shape == (2, 4, 3)
+    assert action[:, 0, 0].tolist() == [1.0, 2.0]
+    assert [item["image_shape"] for item in captured] == [(1, 3, 16, 16), (1, 3, 16, 16)]
+    assert [item["proprio_shape"] for item in captured] == [(1, 2), (1, 2)]
+    assert [item["prompt"] for item in captured] == [
+        cfg.prompt_template.format(task="task 0"),
+        cfg.prompt_template.format(task="task 1"),
+    ]
+
+
+class CoreWithFrozenComponents(FakeFastWAMCore):
+    """Fake core mirroring the real one: frozen VAE / text encoder held as
+    *unregistered* attributes (via `object.__setattr__`) so they are excluded from
+    `state_dict()` and the saved checkpoint, but still moved by the `_apply` override."""
+
+    def __init__(self):
+        super().__init__()
+        object.__setattr__(self, "vae", nn.Linear(2, 2))
+        object.__setattr__(self, "text_encoder", nn.Linear(2, 2))
+        self.vae.requires_grad_(False)
+        self.text_encoder.requires_grad_(False)
+
+    def _apply(self, fn, *args, **kwargs):
+        super()._apply(fn, *args, **kwargs)
+        self.vae._apply(fn)
+        self.text_encoder._apply(fn)
+        return self
+
+
+def test_from_pretrained_uses_base_loader_and_skips_wan_backbone(monkeypatch, tmp_path):
+    cfg = FastWAMConfig(
+        action_dim=3,
+        proprio_dim=2,
+        action_horizon=4,
+        n_action_steps=2,
+        num_video_frames=5,
+        action_video_freq_ratio=1,
+        base_model_id=None,
+    )
+
+    def build_core(self, config):
+        core = CoreWithFrozenComponents()
+        with torch.no_grad():
+            core.dit.weight.fill_(0.5)
+        return core
+
+    monkeypatch.setattr(FastWAMPolicy, "_build_core_model", build_core)
+
+    reference = FastWAMPolicy(cfg)
+    with torch.no_grad():
+        reference.model.dit.weight.fill_(1.25)  # a distinctive, trained-looking weight
+    reference.save_pretrained(tmp_path)
+
+    # Building from Wan2.2 must never happen on a checkpoint load.
+    def fail_if_wan_pretrained_is_loaded(*args, **kwargs):
+        raise AssertionError("from_pretrained must not initialize or download the Wan2.2 backbone")
+
+    monkeypatch.setattr(
+        "lerobot.policies.fastwam.wan.modular.FastWAM.from_wan22_pretrained",
+        fail_if_wan_pretrained_is_loaded,
+    )
+
+    policy = FastWAMPolicy.from_pretrained(tmp_path)
+
+    assert isinstance(policy.model, CoreWithFrozenComponents)
+    # The bundled checkpoint weights overwrote the freshly built (0.5) DiT weights.
+    assert torch.allclose(policy.model.dit.weight, torch.full_like(policy.model.dit.weight, 1.25))
+
+
+def test_save_pretrained_excludes_frozen_components(monkeypatch, tmp_path):
+    cfg = FastWAMConfig(
+        action_dim=3,
+        proprio_dim=2,
+        action_horizon=4,
+        n_action_steps=2,
+        num_video_frames=5,
+        action_video_freq_ratio=1,
+        base_model_id=None,
+    )
+    monkeypatch.setattr(FastWAMPolicy, "_build_core_model", lambda self, config: CoreWithFrozenComponents())
+    policy = FastWAMPolicy(cfg)
+
+    save_dir = tmp_path / "saved"
+    policy.save_pretrained(save_dir)
+
+    assert (save_dir / "model.safetensors").is_file()
+    # No Wan sidecar files either: the frozen backbone comes from the diffusers repo.
+    assert not (save_dir / "Wan2.2_VAE.safetensors").exists()
+    assert not (save_dir / "google").exists()
+
+    with safe_open(save_dir / "model.safetensors", framework="pt") as f:
+        keys = set(f.keys())
+    # Lean checkpoint: only the trainable DiT is saved; the frozen VAE / UMT5 text
+    # encoder are excluded (loaded from the diffusers/transformers repos at init).
+    assert any(key.startswith("model.dit.") for key in keys)
+    assert not any(key.startswith("model.vae.") for key in keys)
+    assert not any(key.startswith("model.text_encoder.") for key in keys)
+
+
+def test_frozen_components_excluded_from_params_but_follow_device_moves(monkeypatch):
+    cfg = FastWAMConfig(
+        action_dim=3,
+        proprio_dim=2,
+        action_horizon=4,
+        n_action_steps=2,
+        num_video_frames=5,
+        action_video_freq_ratio=1,
+        base_model_id=None,
+    )
+    monkeypatch.setattr(FastWAMPolicy, "_build_core_model", lambda self, config: CoreWithFrozenComponents())
+    policy = FastWAMPolicy(cfg)
+
+    # Unregistered: excluded from state_dict and from the optimizer's parameter set.
+    sd = policy.state_dict()
+    assert not any(k.startswith("model.vae.") or k.startswith("model.text_encoder.") for k in sd)
+    param_names = [n for n, _ in policy.named_parameters()]
+    assert not any("vae" in n or "text_encoder" in n for n in param_names)
+
+    # ...but the `_apply` override still carries them through `.to()` (dtype stands in
+    # for device on a CPU box), so they never strand off the rest of the model.
+    policy.to(torch.float64)
+    assert policy.model.dit.weight.dtype == torch.float64  # registered
+    assert policy.model.vae.weight.dtype == torch.float64  # unregistered, moved via _apply
+    assert policy.model.text_encoder.weight.dtype == torch.float64
+
+
+def test_pretrained_config_round_trips_fastwam_features(tmp_path):
+    cfg = FastWAMConfig(action_dim=7, proprio_dim=8, image_size=(224, 448), base_model_id=None)
+    cfg.save_pretrained(tmp_path)
+
+    loaded = PreTrainedConfig.from_pretrained(tmp_path)
+
+    assert loaded.type == "fastwam"
+    assert loaded.image_features["observation.images.image"].type == FeatureType.VISUAL
+    assert loaded.action_feature.shape == (7,)
+    assert loaded.robot_state_feature.shape == (8,)
+
+
+def test_vae_adapter_empty_build_encode_decode_shapes():
+    """Offline glue check of the diffusers-backed VAE adapter (random weights).
+
+    Validates the encode/decode contract — 48 latent channels, 16x spatial / 4x
+    temporal compression, list-or-batch input, scaling round-trip — without any
+    weight download. (Numerical fidelity vs the original Wan VAE is a separate,
+    GPU + real-weights verification step.)
+    """
+    pytest.importorskip("diffusers")
+    from diffusers import AutoencoderKLWan
+
+    from lerobot.policies.fastwam.wan import WanVideoVAE38
+
+    # Production always loads a real pretrained VAE from the diffusers repo; here we
+    # build the same architecture with random weights and dummy standardization stats
+    # to exercise the adapter's shape/scaling contract offline (fidelity is checked
+    # separately, with real weights, on GPU).
+    arch = {
+        "base_dim": 160,
+        "decoder_base_dim": 256,
+        "z_dim": 48,
+        "dim_mult": [1, 2, 4, 4],
+        "num_res_blocks": 2,
+        "attn_scales": [],
+        "temporal_downsample": [False, True, True],
+        "dropout": 0.0,
+        "is_residual": True,
+        "in_channels": 12,
+        "out_channels": 12,
+        "patch_size": 2,
+        "scale_factor_spatial": 16,
+        "scale_factor_temporal": 4,
+        "clip_output": False,
+        "latents_mean": [0.0] * 48,
+        "latents_std": [1.0] * 48,
+    }
+    raw = AutoencoderKLWan.from_config(arch)
+    vae = WanVideoVAE38(dtype=torch.float32, device="cpu", pretrained=raw)
+    assert vae.z_dim == 48
+    assert vae.upsampling_factor == 16
+    assert vae.temporal_downsample_factor == 4
+
+    video = torch.rand(1, 3, 5, 32, 32) * 2 - 1  # [B,C,T,H,W] in [-1,1]
+    latents = vae.encode(video)
+    assert latents.shape == (1, 48, 2, 2, 2)  # T'=(5-1)//4+1, H'=W'=32//16
+
+    decoded = vae.decode(latents)
+    assert decoded.shape[0] == 1 and decoded.shape[1] == 3 and decoded.shape[-2:] == (32, 32)
+    assert decoded.min() >= -1.0 and decoded.max() <= 1.0
+
+    # list input is accepted and equals the batched path
+    assert torch.equal(vae.encode([video[0]]), latents)

tests/policies/molmoact2/test_molmoact2.py

@@ -1,5 +1,3 @@
-#!/usr/bin/env python
-
 # Copyright 2026 The Allen Institute for Artificial Intelligence and The HuggingFace Inc. team. All rights reserved.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -35,24 +33,27 @@ pytest.importorskip("scipy")
 from lerobot.configs import FeatureType, NormalizationMode, PolicyFeature
 from lerobot.policies import get_policy_class, make_policy_config
 from lerobot.policies.molmoact2 import (
-    configuration_molmoact2 as molmoact2_config,
     modeling_molmoact2 as molmoact2_modeling,
     processor_molmoact2 as molmoact2_processor,
 )
-from lerobot.policies.molmoact2.configuration_molmoact2 import (
-    MolmoAct2Config,
-    MolmoAct2CosineDecayWithWarmupSchedulerConfig,
-    infer_molmoact2_max_sequence_length,
+from lerobot.policies.molmoact2.configuration_molmoact2 import MolmoAct2Config
+from lerobot.policies.molmoact2.modeling_molmoact2 import (
+    MolmoAct2Policy,
+    _apply_action_chunk_padding_mask,
+    _apply_action_dim_padding_mask,
+    _combine_rollout_seeds,
 )
-from lerobot.policies.molmoact2.modeling_molmoact2 import MolmoAct2Policy
 from lerobot.policies.molmoact2.processor_molmoact2 import (
+    MolmoAct2ActionFrameTransformStep,
     MolmoAct2ClampNormalizedProcessorStep,
     MolmoAct2MaskedNormalizerProcessorStep,
     MolmoAct2MaskedUnnormalizerProcessorStep,
     MolmoAct2PackInputsProcessorStep,
+    MolmoAct2StateFrameTransformStep,
     _add_gripper_masks_to_stats,
     _build_discrete_state_string,
     _normalize_question_text,
+    infer_molmoact2_max_sequence_length,
     make_molmoact2_pre_post_processors,
 )
 from lerobot.policies.rtc.configuration_rtc import RTCConfig
@@ -71,34 +72,38 @@ def test_molmoact2_policy_registration():
     assert cfg.per_episode_seed is False
     assert cfg.eval_seed is None
     assert cfg.normalize_language is True
-    assert cfg.get_scheduler_preset().num_decay_steps is None
+    assert cfg.get_scheduler_preset().num_decay_steps == 100_000
     assert cfg.action_delta_indices == list(range(cfg.chunk_size))
     assert get_policy_class("molmoact2") is MolmoAct2Policy
 
 
 def test_molmoact2_checkpoint_download_ignores_remote_python(monkeypatch):
+    import huggingface_hub
+
     download_kwargs = {}
 
     def fake_snapshot_download(**kwargs):
         download_kwargs.update(kwargs)
         return "/tmp/downloaded-molmoact2"
 
-    monkeypatch.setattr(molmoact2_config, "snapshot_download", fake_snapshot_download)
+    monkeypatch.setattr(huggingface_hub, "snapshot_download", fake_snapshot_download)
 
-    checkpoint_location = molmoact2_config._resolve_checkpoint_location("allenai/MolmoAct2")
+    checkpoint_location = molmoact2_modeling._resolve_checkpoint_location("allenai/MolmoAct2")
 
     assert checkpoint_location == "/tmp/downloaded-molmoact2"
     assert download_kwargs["ignore_patterns"] == ["*.py", "*.pyc", "__pycache__/*"]
 
 
-def test_molmoact2_scheduler_decay_steps_auto_match_training_steps():
+def test_molmoact2_scheduler_auto_scales_to_training_steps():
+    from lerobot.optim import CosineDecayWithWarmupSchedulerConfig
+
     param = torch.nn.Parameter(torch.ones(()))
     optimizer = torch.optim.AdamW([param], lr=0.001)
-    config = MolmoAct2CosineDecayWithWarmupSchedulerConfig(
+    config = CosineDecayWithWarmupSchedulerConfig(
         peak_lr=0.01,
         decay_lr=0.001,
         num_warmup_steps=10,
-        num_decay_steps=None,
+        num_decay_steps=100_000,
     )
 
     scheduler = config.build(optimizer, num_training_steps=100)
@@ -123,9 +128,7 @@ def test_molmoact2_rollout_generator_uses_eval_seed_per_task():
         batch_size=3,
         device=torch.device("cpu"),
     )
-    expected_first = torch.Generator().manual_seed(
-        MolmoAct2Policy._combine_rollout_seeds(first_seed=1000, batch_size=3)
-    )
+    expected_first = torch.Generator().manual_seed(_combine_rollout_seeds(first_seed=1000, batch_size=3))
     assert torch.allclose(torch.rand(4, generator=first), torch.rand(4, generator=expected_first))
 
     policy.reset()
@@ -134,9 +137,7 @@ def test_molmoact2_rollout_generator_uses_eval_seed_per_task():
         batch_size=3,
         device=torch.device("cpu"),
     )
-    expected_second = torch.Generator().manual_seed(
-        MolmoAct2Policy._combine_rollout_seeds(first_seed=1003, batch_size=3)
-    )
+    expected_second = torch.Generator().manual_seed(_combine_rollout_seeds(first_seed=1003, batch_size=3))
     assert torch.allclose(torch.rand(4, generator=second), torch.rand(4, generator=expected_second))
 
     policy.reset()
@@ -145,9 +146,7 @@ def test_molmoact2_rollout_generator_uses_eval_seed_per_task():
         batch_size=3,
         device=torch.device("cpu"),
     )
-    expected_new_task = torch.Generator().manual_seed(
-        MolmoAct2Policy._combine_rollout_seeds(first_seed=1000, batch_size=3)
-    )
+    expected_new_task = torch.Generator().manual_seed(_combine_rollout_seeds(first_seed=1000, batch_size=3))
     assert torch.allclose(torch.rand(4, generator=new_task), torch.rand(4, generator=expected_new_task))
 
 
@@ -537,36 +536,26 @@ def test_train_action_expert_only_requires_continuous_action_mode():
 
 
 def test_molmoact2_sequence_length_is_inferred_from_fixed_token_budget():
-    cfg = MolmoAct2Config(
-        action_mode="both",
-        chunk_size=10,
-        n_action_steps=10,
-        image_keys=["observation.images.image", "observation.images.wrist_image"],
-        input_features={OBS_STATE: PolicyFeature(type=FeatureType.STATE, shape=(8,))},
-        output_features={ACTION: PolicyFeature(type=FeatureType.ACTION, shape=(7,))},
-    )
-
-    assert cfg.max_sequence_length is None
-    assert cfg.inferred_max_sequence_length() == 640
-    assert cfg.inferred_max_sequence_length(include_discrete_action=False) == 576
     assert (
         infer_molmoact2_max_sequence_length(
-            num_images=2,
-            state_dim=8,
-            action_dim=7,
-            action_horizon=30,
-            include_discrete_action=True,
+            num_images=2, state_dim=8, action_dim=7, action_horizon=10, include_discrete_action=True
+        )
+        == 640
+    )
+    assert (
+        infer_molmoact2_max_sequence_length(
+            num_images=2, state_dim=8, action_dim=7, action_horizon=10, include_discrete_action=False
+        )
+        == 576
+    )
+    assert (
+        infer_molmoact2_max_sequence_length(
+            num_images=2, state_dim=8, action_dim=7, action_horizon=30, include_discrete_action=True
         )
         == 768
     )
 
 
-def test_molmoact2_sequence_length_override_is_preserved():
-    cfg = MolmoAct2Config(max_sequence_length=1024)
-
-    assert cfg.inferred_max_sequence_length(num_images=2, state_dim=8, action_dim=7) == 1024
-
-
 def test_train_action_expert_only_freezes_non_action_expert_params():
     class DummyBackbone(torch.nn.Module):
         def __init__(self):
@@ -939,6 +928,39 @@ def test_question_normalization_matches_release_prompt_style():
     )
 
 
+def test_joint_frame_transform_round_trip():
+    signs = [1.0, -1.0, 1.0, 1.0, 1.0, 1.0]
+    offsets = [0.0, 90.0, 90.0, 0.0, 0.0, 0.0]
+    original_state = torch.tensor([[10.0, -90.0, -120.0, 30.0, 0.0, -45.0]])
+
+    state_step = MolmoAct2StateFrameTransformStep(joint_signs=signs, joint_offsets=offsets)
+    action_step = MolmoAct2ActionFrameTransformStep(joint_signs=signs, joint_offsets=offsets)
+
+    transition = {
+        TransitionKey.OBSERVATION: {OBS_STATE: original_state.clone()},
+    }
+    transformed = state_step(transition)
+    model_state = transformed[TransitionKey.OBSERVATION][OBS_STATE]
+
+    action_transition = {TransitionKey.ACTION: model_state.clone()}
+    recovered = action_step(action_transition)
+    recovered_state = recovered[TransitionKey.ACTION]
+
+    assert torch.allclose(recovered_state, original_state)
+
+
+def test_joint_frame_transform_noop_when_none():
+    state_step = MolmoAct2StateFrameTransformStep(joint_signs=None, joint_offsets=None)
+    action_step = MolmoAct2ActionFrameTransformStep(joint_signs=None, joint_offsets=None)
+    state = torch.tensor([[10.0, -90.0, -120.0]])
+
+    state_transition = {TransitionKey.OBSERVATION: {OBS_STATE: state}}
+    assert state_step(state_transition) is state_transition
+
+    action_transition = {TransitionKey.ACTION: state}
+    assert action_step(action_transition) is action_transition
+
+
 def test_action_padding_marks_only_real_dimensions():
     step = object.__new__(MolmoAct2PackInputsProcessorStep)
     step.max_action_dim = 32
@@ -963,7 +985,7 @@ def test_action_dim_padding_loss_reduces_like_old_trainer():
         ]
     )
 
-    reduced = MolmoAct2Policy._apply_action_dim_padding_mask(loss, action_dim_is_pad)
+    reduced = _apply_action_dim_padding_mask(loss, action_dim_is_pad)
 
     expected = torch.stack(
         [
@@ -979,7 +1001,7 @@ def test_action_chunk_padding_keeps_old_mean_denominator():
     loss = torch.ones(1, 2, 4, 3)
     action_horizon_is_pad = torch.tensor([[False, False, True, True]])
 
-    masked = MolmoAct2Policy._apply_action_chunk_padding_mask(loss, action_horizon_is_pad)
+    masked = _apply_action_chunk_padding_mask(loss, action_horizon_is_pad)
 
     assert masked.mean().item() == 0.5
 

tests/policies/vla_jepa/conftest.py

@@ -8,7 +8,6 @@ from types import SimpleNamespace
 import numpy as np
 import pytest
 import torch
-from PIL import Image
 from torch import Tensor, nn
 
 from lerobot.configs.types import FeatureType, PolicyFeature
@@ -191,7 +190,7 @@ class _FakeQwenInterface(nn.Module):
 
     def build_inputs(
         self,
-        images: list[list[Image.Image]],
+        images: list[list[Tensor]],
         instructions: list[str],
         action_prompt: str,
         embodied_prompt: str,
@@ -214,12 +213,13 @@ class _FakeQwenInterface(nn.Module):
         }
 
     @staticmethod
-    def tensor_to_pil(image_tensor: Tensor) -> Image.Image:
-        image = image_tensor.detach().cpu()
-        if image.ndim == 3 and image.shape[0] in (1, 3):
-            image = image.permute(1, 2, 0)
-        image = (image.float().clamp(0, 1) * 255).to(torch.uint8).numpy()
-        return Image.fromarray(image)
+    def to_pixel_values(image_tensor: Tensor) -> Tensor:
+        image = image_tensor.detach().float()
+        if image.shape[-3] == 1:
+            repeats = [1] * image.ndim
+            repeats[-3] = 3
+            image = image.repeat(*repeats)
+        return image
 
 
 class _FakeVideoEncoder(nn.Module):
@@ -242,12 +242,14 @@ class _FakeVideoEncoder(nn.Module):
 
 
 class _FakeVideoProcessor:
-    def __call__(self, videos, return_tensors: str) -> dict[str, Tensor]:
+    def __call__(self, videos, return_tensors: str, device=None, **kwargs) -> dict[str, Tensor]:
         assert return_tensors == "pt"
         if isinstance(videos, list):
             pixel_values = torch.stack([torch.as_tensor(v) for v in videos])
         else:
             pixel_values = torch.as_tensor(videos).unsqueeze(0)
+        if device is not None:
+            pixel_values = pixel_values.to(device)
         return {"pixel_values_videos": pixel_values}
 
 

tests/policies/vla_jepa/test_vla_jepa.py

@@ -211,40 +211,42 @@ def test_reset_clears_action_queue(patch_vla_jepa_external_models: None) -> None
 
 
 def test_prepare_model_inputs_training_format(patch_vla_jepa_external_models: None) -> None:
-    from PIL import Image
-
     policy = VLAJEPAPolicy(make_config())
-    examples = policy._prepare_model_inputs(make_train_batch())
+    inputs = policy._prepare_model_inputs(make_train_batch())
 
-    assert len(examples) == BATCH_SIZE
-    for ex in examples:
-        assert set(ex) >= {"image", "video", "lang", "action", "state"}
-        assert len(ex["image"]) == 1 and isinstance(ex["image"][0], Image.Image)
-        assert ex["video"].ndim == 5 and ex["video"].dtype == np.uint8  # [V,T,H,W,C]
-        assert ex["action"].shape == (ACTION_HORIZON, ACTION_DIM)
-        assert ex["state"].shape == (1, STATE_DIM)
+    assert set(inputs) >= {"images", "instructions", "videos", "actions", "state"}
+    # images: per-sample, per-view [C, H, W] float tensors (kept as a list for Qwen messages)
+    assert len(inputs["images"]) == BATCH_SIZE and len(inputs["images"][0]) == 1
+    img = inputs["images"][0][0]
+    assert isinstance(img, torch.Tensor) and img.dtype == torch.float32 and img.ndim == 3
+    assert len(inputs["instructions"]) == BATCH_SIZE
+    # videos: batched [B, V, T, C, H, W] float
+    assert inputs["videos"].ndim == 6 and inputs["videos"].shape[0] == BATCH_SIZE
+    assert inputs["videos"].dtype == torch.float32
+    assert inputs["actions"].shape == (BATCH_SIZE, ACTION_HORIZON, ACTION_DIM)
+    assert inputs["state"].shape == (BATCH_SIZE, 1, STATE_DIM)
 
 
 def test_prepare_model_inputs_inference_omits_action(patch_vla_jepa_external_models: None) -> None:
     policy = VLAJEPAPolicy(make_config())
-    for ex in policy._prepare_model_inputs(make_inference_batch()):
-        assert "action" not in ex
-        assert "image" in ex and "video" in ex and "lang" in ex
+    inputs = policy._prepare_model_inputs(make_inference_batch())
+    assert "actions" not in inputs and "action_is_pad" not in inputs
+    assert {"images", "instructions", "state"} <= set(inputs)
 
 
 def test_prepare_model_inputs_missing_task_uses_default(patch_vla_jepa_external_models: None) -> None:
     policy = VLAJEPAPolicy(make_config())
     batch = make_inference_batch()
     del batch["task"]
-    examples = policy._prepare_model_inputs(batch)
-    assert all(isinstance(ex["lang"], str) and len(ex["lang"]) > 0 for ex in examples)
+    instructions = policy._prepare_model_inputs(batch)["instructions"]
+    assert all(isinstance(s, str) and len(s) > 0 for s in instructions)
 
 
 def test_prepare_model_inputs_string_task_broadcast(patch_vla_jepa_external_models: None) -> None:
     policy = VLAJEPAPolicy(make_config())
     batch = make_inference_batch()
     batch["task"] = "open the drawer"
-    assert all(ex["lang"] == "open the drawer" for ex in policy._prepare_model_inputs(batch))
+    assert policy._prepare_model_inputs(batch)["instructions"] == ["open the drawer"] * BATCH_SIZE
 
 
 def test_prepare_model_inputs_no_state_omitted(patch_vla_jepa_external_models: None) -> None:
@@ -253,7 +255,7 @@ def test_prepare_model_inputs_no_state_omitted(patch_vla_jepa_external_models: N
     policy = VLAJEPAPolicy(make_config())
     batch = make_inference_batch()
     del batch[OBS_STATE]
-    assert all("state" not in ex for ex in policy._prepare_model_inputs(batch))
+    assert "state" not in policy._prepare_model_inputs(batch)
 
 
 # ---------------------------------------------------------------------------
@@ -446,14 +448,14 @@ def test_postprocessor_applied_after_predict_action_chunk(
     """
     from lerobot.policies.vla_jepa.processor_vla_jepa import make_vla_jepa_pre_post_processors
 
-    raw_actions = np.zeros((BATCH_SIZE, ACTION_HORIZON, ACTION_DIM), dtype=np.float32)
+    raw_actions = torch.zeros((BATCH_SIZE, ACTION_HORIZON, ACTION_DIM), dtype=torch.float32)
 
     cfg = make_config()
     cfg.clip_normalized_actions = False
     cfg.binarize_gripper_action = False
     policy = VLAJEPAPolicy(cfg)
     policy.eval()
-    monkeypatch.setattr(policy.model, "predict_action", lambda *a, **kw: raw_actions.copy())
+    monkeypatch.setattr(policy.model, "predict_action", lambda *a, **kw: raw_actions.clone())
 
     dataset_stats = _make_dataset_stats()
     _, postprocessor = make_vla_jepa_pre_post_processors(cfg, dataset_stats)
@@ -564,9 +566,9 @@ def test_single_view_is_duplicated_for_world_model(patch_vla_jepa_external_model
     original_processor = policy.model.video_processor
 
     class _CapturingProcessor:
-        def __call__(self, videos: list, return_tensors: str) -> dict:
+        def __call__(self, videos: list, return_tensors: str, **kwargs) -> dict:
             captured_videos.extend(videos)
-            return original_processor(videos=videos, return_tensors=return_tensors)
+            return original_processor(videos=videos, return_tensors=return_tensors, **kwargs)
 
     policy.model.video_processor = _CapturingProcessor()
     policy.forward(_make_multiview_train_batch(num_views=1))
@@ -587,9 +589,9 @@ def test_excess_views_trimmed_for_world_model(patch_vla_jepa_external_models: No
     original_processor = policy.model.video_processor
 
     class _CapturingProcessor:
-        def __call__(self, videos: list, return_tensors: str) -> dict:
+        def __call__(self, videos: list, return_tensors: str, **kwargs) -> dict:
             captured_videos.extend(videos)
-            return original_processor(videos=videos, return_tensors=return_tensors)
+            return original_processor(videos=videos, return_tensors=return_tensors, **kwargs)
 
     policy.model.video_processor = _CapturingProcessor()
     policy.forward(_make_multiview_train_batch(num_views=3))