feat(vla): extract shared action-time expert embedding block

Add fuse_action_time_embedding to common/vla_utils: the action-expert input
block (project actions + sine-cosine timestep embed + concat + mlp_in/SiLU/
mlp_out) that pi0, pi05, eo1 and smolvla each copy. The nn.Linear layers are
passed in rather than owned by the helper, so adoption does not rename any
checkpoint keys.

Reproduces the pi0/pi05 (time_emb_dtype=timestep.dtype) and smolvla (default
action-emb dtype) conventions byte-for-byte, with an optional apply hook for
gradient checkpointing. eo1's autocast + per-layer dtype casts are left for a
follow-up adoption. No policy is migrated in this (additive-only) PR.

Adds equivalence tests against verbatim pi0 and smolvla inline blocks.

Co-authored-by: Cursor <cursoragent@cursor.com>

docs/source/bring_your_own_policies.mdx

@@ -150,14 +150,14 @@ class MyPolicy(PreTrainedPolicy):
 
 The methods called by the train/eval loops:
 
-| Method                                                            | Used by           | What it does                                                                                                                                                                                                                                         |
-| ----------------------------------------------------------------- | ----------------- | ---------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
-| `reset() -> None`                                                 | `lerobot-eval`    | Clear per-episode state at the start of each episode.                                                                                                                                                                                                |
-| `select_action(batch, **kwargs) -> Tensor`                        | `lerobot-eval`    | Return the next action `(B, action_dim)`. Called every step.                                                                                                                                                                                         |
-| `predict_action_chunk(batch, **kwargs) -> Tensor`                 | the policy itself | Return an action chunk `(B, chunk_size, action_dim)`. Currently abstract on the base class — raise `NotImplementedError` if your policy doesn't chunk.                                                                                               |
-| `forward(batch, reduction="mean") -> tuple[Tensor, dict \| None]` | `lerobot-train`   | Return `(loss, output_dict)`. Accept `reduction="none"` if you want to support per-sample weighting.                                                                                                                                                 |
-| `get_optim_params() -> dict`                                      | the optimizer     | Return `self.parameters()` for simple policies; return a named parameter dict for [multi-optimizer policies](https://github.com/huggingface/lerobot/blob/ecd38c50d7d15b4184cf42649ff1185ee2e11eeb/src/lerobot/policies/sac/modeling_sac.py#L61-L73). |
-| `update() -> None` _(optional)_                                   | `lerobot-train`   | Called after each optimizer step _if defined_. Use for EMA, target nets, replay buffers (TDMPC uses this).                                                                                                                                           |
+| Method                                                            | Used by           | What it does                                                                                                                                                                                                                                                                                 |
+| ----------------------------------------------------------------- | ----------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
+| `reset() -> None`                                                 | `lerobot-eval`    | Clear per-episode state at the start of each episode.                                                                                                                                                                                                                                        |
+| `select_action(batch, **kwargs) -> Tensor`                        | `lerobot-eval`    | Return the next action `(B, action_dim)`. Called every step.                                                                                                                                                                                                                                 |
+| `predict_action_chunk(batch, **kwargs) -> Tensor`                 | the policy itself | Return an action chunk `(B, chunk_size, action_dim)`. Currently abstract on the base class — raise `NotImplementedError` if your policy doesn't chunk.                                                                                                                                       |
+| `forward(batch, reduction="mean") -> tuple[Tensor, dict \| None]` | `lerobot-train`   | Return `(loss, output_dict)`. Accept `reduction="none"` if you want to support per-sample weighting.                                                                                                                                                                                         |
+| `get_optim_params() -> dict`                                      | the optimizer     | Return `self.parameters()` for simple policies; return a named parameter dict for multi-optimizer policies (see `get_optim_params` in [`modeling_act.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/act/modeling_act.py) for a per-group learning-rate example). |
+| `update() -> None` _(optional)_                                   | `lerobot-train`   | Called after each optimizer step _if defined_. Use for EMA, target nets, replay buffers (TDMPC uses this).                                                                                                                                                                                   |
 
 Batches are flat dictionaries keyed by the constants in [`lerobot.utils.constants`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/utils/constants.py): `OBS_STATE` (`observation.state.<motor>`), `OBS_IMAGES` (`observation.images.<camera>`), `OBS_LANGUAGE`, `ACTION`, etc. Reuse the constants — don't invent new prefixes.
 
@@ -295,12 +295,10 @@ The file names are load-bearing: the factory does lazy imports by name, and the
 
 ### Wiring
 
-Four places need to know about your policy. All by name.
+Two places need to know about your policy. All by name.
 
-1. **`policies/__init__.py`** — re-export `MyPolicyConfig` and add it to `__all__`. **Don't** re-export the modeling class; it loads lazily through the factory (so `import lerobot` stays fast).
-2. **`factory.py:get_policy_class`** — add a branch returning `MyPolicy` from a lazy import.
-3. **`factory.py:make_policy_config`** and **`factory.py:make_pre_post_processors`** — same idea, two more branches.
-4. **`templates/lerobot_modelcard_template.md` and the root `README.md`** — the template is what `push_model_to_hub` renders into the model card of every checkpoint trained with your policy: add a one-line description of your policy in the `model_name` branches, map it in `policy_docs` so cards link to your MDX guide, and optionally add an architecture image to `diagrams`. Then add your policy to the models table in the root `README.md`, under the right category, linking to your doc page.
+1. **`policies/__init__.py`** — re-export `MyPolicyConfig` and add it to `__all__`. This import is what registers your policy: `@PreTrainedConfig.register_subclass("my_policy")` runs, and from then on the factory resolves everything by convention. **Don't** re-export the modeling class; it loads lazily through the factory (so `import lerobot` stays fast).
+2. **`templates/lerobot_modelcard_template.md` and the root `README.md`** — the template is what `push_model_to_hub` renders into the model card of every checkpoint trained with your policy: add a one-line description of your policy in the `model_name` branches, map it in `policy_docs` so cards link to your MDX guide, and optionally add an architecture image to `diagrams`. Then add your policy to the models table in the root `README.md`, under the right category, linking to your doc page.
 
 Mirror an existing policy that's structurally similar to yours; the diff is small.
 
@@ -332,6 +330,10 @@ This way:
 
 Add a matching extra to [`pyproject.toml`](https://github.com/huggingface/lerobot/blob/main/pyproject.toml) `[project.optional-dependencies]` and include it in the `all` extra so `pip install 'lerobot[all]'` keeps installing everything.
 
+### Avoid copying a modeling file — subclass it
+
+If your policy needs to modify a backbone that already exists in `transformers` (custom conditioning, extra inputs, a swapped sub-module), **do not vendor a copy of its `modeling_*.py`**. Instead, subclass the smallest upstream unit and override only what changes. [`pi_gemma.py`](https://github.com/huggingface/lerobot/blob/main/src/lerobot/policies/pi_gemma.py) is the canonical reference: it injects AdaRMS conditioning into PaliGemma/Gemma in ~370 lines by subclassing `GemmaModel`/`PaliGemmaModel` and overriding the decoder-layer forward, instead of forking the ~2,000-line modeling file. Model surgery on a _loaded_ native model is also fine (layer truncation, tokenizer expansion, hidden-state capture — see `evo1/internvl3_embedder.py`, `eo1/modeling_eo1.py`, `groot/groot_n1_7.py` for working examples). Reviewers will ask for this pattern when a PR arrives with a copied modeling file; the only accepted exception is a model that does not exist in `transformers` at all.
+
 ### Benchmarks and a published checkpoint
 
 A new policy is much easier to review — and far more useful — when it ships with a working checkpoint and at least one number you can reproduce.
@@ -367,7 +369,7 @@ If your policy is real-robot-only and no sim benchmark applies, swap the sim eva
 The general expectations are in [`CONTRIBUTING.md`](https://github.com/huggingface/lerobot/blob/main/CONTRIBUTING.md) and the [PR template](https://github.com/huggingface/lerobot/blob/main/.github/PULL_REQUEST_TEMPLATE.md). On top of those, reviewers will look for:
 
 - [ ] `MyPolicy` and `MyPolicyConfig` cover the surface above; `__init_subclass__` accepts the class.
-- [ ] `factory.py` and `policies/__init__.py` are wired (lazy imports for modeling).
+- [ ] `policies/__init__.py` re-exports the config (this registers the policy; the factory resolves modeling/processor by naming convention).
 - [ ] `make_my_policy_pre_post_processors` follows the naming convention.
 - [ ] Optional deps live behind a `[project.optional-dependencies]` extra and the `TYPE_CHECKING + require_package` guard.
 - [ ] `tests/policies/` updated; backward-compat artifact committed & policy-specific tests.

pyproject.toml

@@ -413,8 +413,6 @@ ignore = [
 "__init__.py" = ["F401", "F403", "E402"]
 # E402: conditional-import guards (TYPE_CHECKING / is_package_available) must precede the imports they protect
 "src/lerobot/scripts/convert_dataset_v21_to_v30.py" = ["E402"]
-"src/lerobot/policies/wall_x/**" = ["N801", "N812", "SIM102", "SIM108", "SIM210", "SIM211", "B006", "B007", "SIM118"] # Supprese these as they are coming from original Qwen2_5_vl code TODO(pepijn): refactor original
-
 [tool.ruff.lint.isort]
 combine-as-imports = true
 known-first-party = ["lerobot"]

src/lerobot/configs/policies.py

@@ -205,24 +205,30 @@ class PreTrainedConfig(draccus.ChoiceRegistry, HubMixin, abc.ABC):  # type: igno
                     f"{CONFIG_NAME} not found on the HuggingFace Hub in {model_id}"
                 ) from e
 
-        # HACK: Parse the original config to get the config subclass, so that we can
-        # apply cli overrides.
-        # This is very ugly, ideally we'd like to be able to do that natively with draccus
-        # something like --policy.path (in addition to --policy.type)
-        with draccus.config_type("json"):
-            orig_config = draccus.parse(cls, config_file, args=[])
-
         if config_file is None:
             raise FileNotFoundError(f"{CONFIG_NAME} not found in {model_id}")
 
         with open(config_file) as f:
             config = json.load(f)
 
-        config.pop("type")
+        # Resolve the concrete config subclass from the serialized "type" tag, then parse
+        # the config (with CLI overrides) directly for that class. The "type" key is
+        # stripped because draccus only consumes it when parsing the registry base class.
+        policy_type = config.pop("type", None)
+        if policy_type is None:
+            raise ValueError(f"Missing 'type' field in {CONFIG_NAME} of {model_id}")
+        try:
+            config_cls = cls.get_choice_class(policy_type)
+        except Exception as e:
+            raise ValueError(
+                f"Policy type '{policy_type}' (from {CONFIG_NAME} of {model_id}) is not registered. "
+                f"Available policy types: {cls.get_known_choices()}"
+            ) from e
+
         with tempfile.NamedTemporaryFile("w+", delete=False, suffix=".json") as f:
             json.dump(config, f)
             config_file = f.name
 
         cli_overrides = policy_kwargs.pop("cli_overrides", [])
         with draccus.config_type("json"):
-            return draccus.parse(orig_config.__class__, config_file, args=cli_overrides)
+            return draccus.parse(config_cls, config_file, args=cli_overrides)

src/lerobot/policies/__init__.py

@@ -32,6 +32,7 @@ from .pretrained import PreTrainedPolicy as PreTrainedPolicy
 from .smolvla.configuration_smolvla import SmolVLAConfig as SmolVLAConfig
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
 from .utils import make_robot_action, prepare_observation_for_inference
+from .vla_jepa.configuration_vla_jepa import VLAJEPAConfig as VLAJEPAConfig
 from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig
 from .wall_x.configuration_wall_x import WallXConfig as WallXConfig
 from .xvla.configuration_xvla import XVLAConfig as XVLAConfig
@@ -57,6 +58,7 @@ __all__ = [
     "PI05Config",
     "SmolVLAConfig",
     "TDMPCConfig",
+    "VLAJEPAConfig",
     "VQBeTConfig",
     "WallXConfig",
     "XVLAConfig",

src/lerobot/policies/act/processor_act.py

@@ -18,17 +18,10 @@ from typing import Any
 import torch
 
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_pre_post_processors,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_act import ACTConfig
 
@@ -54,34 +47,4 @@ def make_act_pre_post_processors(
         tuple[PolicyProcessorPipeline[dict[str, Any], dict[str, Any]], PolicyProcessorPipeline[PolicyAction, PolicyAction]]: A tuple containing the
         pre-processor pipeline and the post-processor pipeline.
     """
-
-    input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-            device=config.device,
-        ),
-    ]
-    output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
-    ]
-
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_default_pre_post_processors(config, dataset_stats, normalizer_device=config.device)

src/lerobot/policies/common/flow_matching.py

@@ -0,0 +1,122 @@
+#!/usr/bin/env python
+
+# Copyright 2026 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.
+
+"""Flow-matching sampling primitives shared across policies.
+
+Canonical versions of the beta-distributed timestep sampler and the forward-Euler
+denoising loop (with its real-time-chunking hook) that the openpi-derived policies
+(pi0, pi05, smolvla, eo1) historically each carried a copy of. All functions are
+stateless; adopting them does not affect checkpoints.
+"""
+
+from collections.abc import Callable
+from typing import TYPE_CHECKING
+
+import torch
+from torch import Tensor
+
+if TYPE_CHECKING:
+    from lerobot.policies.rtc.modeling_rtc import RTCProcessor
+
+
+def sample_beta(alpha: float, beta: float, bsize: int, device) -> Tensor:  # see openpi (exact copy)
+    # Beta sampling uses _sample_dirichlet which isn't implemented for MPS, so sample on CPU
+    alpha_t = torch.tensor(alpha, dtype=torch.float32)
+    beta_t = torch.tensor(beta, dtype=torch.float32)
+    dist = torch.distributions.Beta(alpha_t, beta_t)
+    return dist.sample((bsize,)).to(device)
+
+
+def sample_noise(shape, device) -> Tensor:
+    """Standard-normal float32 noise, the flow-matching x_1 sample."""
+    return torch.normal(
+        mean=0.0,
+        std=1.0,
+        size=shape,
+        dtype=torch.float32,
+        device=device,
+    )
+
+
+def sample_time_beta(bsize: int, device, *, alpha: float, beta: float, scale: float, offset: float) -> Tensor:
+    """Beta-distributed flow-matching timesteps: ``Beta(alpha, beta) * scale + offset`` (openpi convention)."""
+    time_beta = sample_beta(alpha, beta, bsize, device)
+    time = time_beta * scale + offset
+    return time.to(dtype=torch.float32, device=device)
+
+
+def euler_integrate(
+    denoise_fn: Callable[[Tensor, Tensor], Tensor],
+    noise: Tensor,
+    num_steps: int,
+    *,
+    rtc_processor: "RTCProcessor | None" = None,
+    rtc_enabled: bool = False,
+    inference_delay: int | None = None,
+    prev_chunk_left_over: Tensor | None = None,
+    execution_horizon: int | None = None,
+) -> Tensor:
+    """Forward-Euler integration of a velocity field from t=1 (noise) to t=0 (actions).
+
+    This is the openpi sampling loop: ``dt = -1/num_steps``, ``time = 1.0 + step*dt``,
+    ``x_t <- x_t + dt * v_t``, with the optional real-time-chunking (RTC) guidance hook
+    wrapping the velocity computation and debug tracking after each step.
+
+    Args:
+        denoise_fn: Computes the velocity ``v_t`` from ``(x_t, time_tensor)`` where
+            ``time_tensor`` is a float32 tensor of shape ``(batch_size,)``. The returned
+            velocity must have the same shape and dtype as ``x_t``.
+        noise: Initial sample ``x_1`` of shape ``(batch_size, ...)``.
+        num_steps: Number of Euler steps.
+        rtc_processor: Optional RTC processor. Debug tracking fires whenever it is set and
+            has debugging enabled, even if RTC guidance itself is disabled (this mirrors
+            the historical per-policy loops).
+        rtc_enabled: Whether to route the velocity computation through
+            ``rtc_processor.denoise_step`` (requires ``rtc_processor``).
+        inference_delay: RTC guidance parameter, forwarded verbatim.
+        prev_chunk_left_over: RTC guidance parameter, forwarded verbatim.
+        execution_horizon: RTC guidance parameter, forwarded verbatim.
+    """
+    bsize = noise.shape[0]
+    device = noise.device
+
+    dt = -1.0 / num_steps
+    x_t = noise
+    for step in range(num_steps):
+        time = 1.0 + step * dt
+        time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
+
+        def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
+            return denoise_fn(input_x_t, current_timestep)
+
+        if rtc_enabled:
+            v_t = rtc_processor.denoise_step(
+                x_t=x_t,
+                prev_chunk_left_over=prev_chunk_left_over,
+                inference_delay=inference_delay,
+                time=time,
+                original_denoise_step_partial=denoise_step_partial_call,
+                execution_horizon=execution_horizon,
+            )
+        else:
+            v_t = denoise_step_partial_call(x_t)
+
+        x_t = x_t + dt * v_t
+
+        if rtc_processor is not None and rtc_processor.is_debug_enabled():
+            rtc_processor.track(time=time, x_t=x_t, v_t=v_t)
+
+    return x_t

src/lerobot/policies/common/vla_utils.py

@@ -0,0 +1,302 @@
+#!/usr/bin/env python
+
+# Copyright 2026 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.
+
+"""Helpers shared by the openpi-derived VLA policies (pi0, pi05, pi0_fast, smolvla, eo1, xvla).
+
+These are the canonical versions of functions that historically were copy-pasted per
+policy. They are pure (no parameters, no module state), so importing them from here
+instead of a policy-local copy has no effect on checkpoints.
+"""
+
+import math
+from collections.abc import Callable
+from typing import TYPE_CHECKING
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F  # noqa: N812
+from torch import Tensor
+
+from lerobot.utils.constants import OPENPI_ATTENTION_MASK_VALUE
+from lerobot.utils.device_utils import get_safe_dtype
+from lerobot.utils.import_utils import _transformers_available, require_package
+
+if TYPE_CHECKING or _transformers_available:
+    from transformers import DynamicCache
+else:
+    DynamicCache = None
+
+
+def create_sinusoidal_pos_embedding(  # see openpi `create_sinusoidal_pos_embedding` (exact copy)
+    time: torch.Tensor, dimension: int, min_period: float, max_period: float, device="cpu"
+) -> Tensor:
+    """Computes sine-cosine positional embedding vectors for scalar positions."""
+    if dimension % 2 != 0:
+        raise ValueError(f"dimension ({dimension}) must be divisible by 2")
+
+    if time.ndim != 1:
+        raise ValueError("The time tensor is expected to be of shape `(batch_size, )`.")
+
+    dtype = get_safe_dtype(torch.float64, device.type)
+    fraction = torch.linspace(0.0, 1.0, dimension // 2, dtype=dtype, device=device)
+    period = min_period * (max_period / min_period) ** fraction
+
+    # Compute the outer product
+    scaling_factor = 1.0 / period * 2 * math.pi
+    sin_input = scaling_factor[None, :] * time[:, None]
+    return torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
+
+
+def fuse_action_time_embedding(
+    noisy_actions: Tensor,
+    timestep: Tensor,
+    *,
+    action_in_proj: nn.Module,
+    action_time_mlp_in: nn.Module,
+    action_time_mlp_out: nn.Module,
+    min_period: float,
+    max_period: float,
+    time_emb_dtype: torch.dtype | None = None,
+    apply: Callable[[Callable, Tensor], Tensor] | None = None,
+) -> Tensor:
+    """Fuse noisy actions and a diffusion timestep into the action-expert input embedding.
+
+    This is the block copy-pasted across the openpi action-expert policies (pi0, pi05,
+    eo1, smolvla): project actions, add a sine-cosine timestep embedding, concatenate, and
+    run ``mlp_in -> SiLU -> mlp_out``. The ``nn.Linear`` layers are passed in (not owned by
+    this helper) so adopting it does not rename any checkpoint keys.
+
+    Args:
+        noisy_actions: ``(batch, horizon, action_dim)`` noised action chunk.
+        timestep: ``(batch,)`` diffusion timestep.
+        action_in_proj: ``Linear(action_dim, width)``.
+        action_time_mlp_in: ``Linear(2*width, width)``.
+        action_time_mlp_out: ``Linear(width, width)``.
+        min_period / max_period: sine-cosine embedding periods.
+        time_emb_dtype: dtype to cast the time embedding to. ``None`` (default, the
+            smolvla/eo1 convention) uses the projected action dtype; pass ``timestep.dtype``
+            for the pi0/pi05 convention.
+        apply: optional wrapper ``apply(fn, arg) -> fn(arg)`` used to route the two
+            sub-computations through gradient checkpointing (pi0/eo1). Defaults to a direct
+            call (smolvla).
+    """
+    if apply is None:
+
+        def apply(fn, arg):
+            return fn(arg)
+
+    action_emb = apply(action_in_proj, noisy_actions)
+
+    time_emb = create_sinusoidal_pos_embedding(
+        timestep,
+        action_in_proj.out_features,
+        min_period=min_period,
+        max_period=max_period,
+        device=timestep.device,
+    )
+    time_emb = time_emb.type(dtype=time_emb_dtype if time_emb_dtype is not None else action_emb.dtype)
+    time_emb = time_emb[:, None, :].expand_as(action_emb)
+    action_time_emb = torch.cat([action_emb, time_emb], dim=2)
+
+    def _mlp(x):
+        return action_time_mlp_out(F.silu(action_time_mlp_in(x)))
+
+    return apply(_mlp, action_time_emb)
+
+
+def make_att_2d_masks(pad_masks: Tensor, att_masks: Tensor) -> Tensor:  # see openpi (exact copy)
+    """Copied from big_vision.
+
+    Tokens can attend to valid inputs tokens which have a cumulative mask_ar
+    smaller or equal to theirs. This way `mask_ar` int[B, N] can be used to
+    setup several types of attention, for example:
+
+      [[1 1 1 1 1 1]]: pure causal attention.
+
+      [[0 0 0 1 1 1]]: prefix-lm attention. The first 3 tokens can attend between
+          themselves and the last 3 tokens have a causal attention. The first
+          entry could also be a 1 without changing behaviour.
+
+      [[1 0 1 0 1 0 0 1 0 0]]: causal attention between 4 blocks. Tokens of a
+          block can attend all previous blocks and all tokens on the same block.
+
+    Args:
+      input_mask: bool[B, N] true if its part of the input, false if padding.
+      mask_ar: int32[B, N] mask that's 1 where previous tokens cannot depend on
+        it and 0 where it shares the same attention mask as the previous token.
+    """
+    if att_masks.ndim != 2:
+        raise ValueError(att_masks.ndim)
+    if pad_masks.ndim != 2:
+        raise ValueError(pad_masks.ndim)
+
+    cumsum = torch.cumsum(att_masks, dim=1)
+    att_2d_masks = cumsum[:, None, :] <= cumsum[:, :, None]
+    pad_2d_masks = pad_masks[:, None, :] * pad_masks[:, :, None]
+    return att_2d_masks & pad_2d_masks
+
+
+def prepare_attention_masks_4d(att_2d_masks: Tensor, dtype: torch.dtype | None = None) -> Tensor:
+    """Expand boolean 2D attention masks to the additive 4D layout expected by transformers.
+
+    Valid positions become 0.0 and masked positions the large negative openpi constant.
+    """
+    att_2d_masks_4d = att_2d_masks[:, None, :, :]
+    result = torch.where(att_2d_masks_4d, 0.0, OPENPI_ATTENTION_MASK_VALUE)
+    if dtype is not None:
+        result = result.to(dtype=dtype)
+    return result
+
+
+def clone_past_key_values(past_key_values):
+    """Clone the DynamicCache returned by prefix prefill for compiled denoising."""
+    if DynamicCache is None:
+        require_package("transformers", extra="transformers-dep")
+
+    return DynamicCache(
+        tuple(
+            (keys.clone(), values.clone(), sliding_window) for keys, values, sliding_window in past_key_values
+        )
+    )
+
+
+def pad_vector(vector: Tensor, new_dim: int, *, truncate: bool = False) -> Tensor:
+    """Pad the last dimension of a vector to new_dim with zeros.
+
+    Can be (batch_size x sequence_length x features_dimension)
+    or (batch_size x features_dimension)
+
+    With ``truncate=False`` (openpi behavior), vectors whose last dimension is already
+    >= new_dim are returned unchanged. With ``truncate=True`` (xVLA behavior), the last
+    dimension is truncated to exactly ``new_dim`` (which may be 0).
+    """
+    if vector.shape[-1] == new_dim:
+        return vector
+    if not truncate:
+        if vector.shape[-1] >= new_dim:
+            return vector
+        return F.pad(vector, (0, new_dim - vector.shape[-1]))
+    shape = list(vector.shape)
+    current_dim = shape[-1]
+    shape[-1] = new_dim
+    new_vector = vector.new_zeros(*shape)
+    length = min(current_dim, new_dim)
+    new_vector[..., :length] = vector[..., :length]
+    return new_vector
+
+
+def resize_with_pad_torch(  # see openpi `resize_with_pad_torch` (exact copy)
+    images: torch.Tensor,
+    height: int,
+    width: int,
+    mode: str = "bilinear",
+) -> torch.Tensor:
+    """PyTorch version of resize_with_pad. Resizes an image to a target height and width without distortion
+    by padding with black. If the image is float32, it must be in the range [-1, 1].
+
+    Padding is centered (openpi convention). For the top-left-padding variant used by
+    smolvla/xvla, see :func:`resize_with_pad`.
+
+    Args:
+        images: Tensor of shape [*b, h, w, c] or [*b, c, h, w]
+        height: Target height
+        width: Target width
+        mode: Interpolation mode ('bilinear', 'nearest', etc.)
+
+    Returns:
+        Resized and padded tensor with same shape format as input
+    """
+    # Check if input is in channels-last format [*b, h, w, c] or channels-first [*b, c, h, w]
+    if images.shape[-1] <= 4:  # Assume channels-last format
+        channels_last = True
+        if images.dim() == 3:
+            images = images.unsqueeze(0)  # Add batch dimension
+        images = images.permute(0, 3, 1, 2)  # [b, h, w, c] -> [b, c, h, w]
+    else:
+        channels_last = False
+        if images.dim() == 3:
+            images = images.unsqueeze(0)  # Add batch dimension
+
+    batch_size, channels, cur_height, cur_width = images.shape
+
+    # Calculate resize ratio
+    ratio = max(cur_width / width, cur_height / height)
+    resized_height = int(cur_height / ratio)
+    resized_width = int(cur_width / ratio)
+
+    # Resize
+    resized_images = F.interpolate(
+        images,
+        size=(resized_height, resized_width),
+        mode=mode,
+        align_corners=False if mode == "bilinear" else None,
+    )
+
+    # Handle dtype-specific clipping
+    if images.dtype == torch.uint8:
+        resized_images = torch.round(resized_images).clamp(0, 255).to(torch.uint8)
+    elif images.dtype == torch.float32:
+        resized_images = resized_images.clamp(0.0, 1.0)
+    else:
+        raise ValueError(f"Unsupported image dtype: {images.dtype}")
+
+    # Calculate padding
+    pad_h0, remainder_h = divmod(height - resized_height, 2)
+    pad_h1 = pad_h0 + remainder_h
+    pad_w0, remainder_w = divmod(width - resized_width, 2)
+    pad_w1 = pad_w0 + remainder_w
+
+    # Pad
+    constant_value = 0 if images.dtype == torch.uint8 else 0.0
+    padded_images = F.pad(
+        resized_images,
+        (pad_w0, pad_w1, pad_h0, pad_h1),  # left, right, top, bottom
+        mode="constant",
+        value=constant_value,
+    )
+
+    # Convert back to original format if needed
+    if channels_last:
+        padded_images = padded_images.permute(0, 2, 3, 1)  # [b, c, h, w] -> [b, h, w, c]
+
+    return padded_images
+
+
+def resize_with_pad(img: torch.Tensor, height: int, width: int, *, pad_value: float) -> torch.Tensor:
+    """Resize a (b, c, h, w) image without distortion, padding on the LEFT and TOP.
+
+    This is the smolvla/xvla convention. For the centered-padding openpi variant, see
+    :func:`resize_with_pad_torch`. ``pad_value`` is keyword-only on purpose: callers
+    historically used different values (0, -1) and must state their choice explicitly.
+    """
+    if img.ndim != 4:
+        raise ValueError(f"(b,c,h,w) expected, but got {img.shape}")
+
+    current_height, current_width = img.shape[2:]
+    if current_height == height and current_width == width:
+        return img
+
+    ratio = max(current_width / width, current_height / height)
+    resized_height = int(current_height / ratio)
+    resized_width = int(current_width / ratio)
+    resized_img = F.interpolate(
+        img, size=(resized_height, resized_width), mode="bilinear", align_corners=False
+    )
+
+    pad_height = max(0, height - resized_height)
+    pad_width = max(0, width - resized_width)
+    padded_img = F.pad(resized_img, (pad_width, 0, pad_height, 0), value=pad_value)
+    return padded_img

src/lerobot/policies/diffusion/processor_diffusion.py

@@ -19,17 +19,10 @@ from typing import Any
 import torch
 
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_pre_post_processors,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_diffusion import DiffusionConfig
 
@@ -63,32 +56,4 @@ def make_diffusion_pre_post_processors(
     Returns:
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
-
-    input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-    ]
-    output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
-    ]
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_default_pre_post_processors(config, dataset_stats)

src/lerobot/policies/eo1/processor_eo1.py

@@ -23,24 +23,16 @@ import torch
 
 from lerobot.configs.types import FeatureType, PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
     ComplementaryDataProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStep,
     ProcessorStepRegistry,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
-from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
 from lerobot.types import TransitionKey
-from lerobot.utils.constants import (
-    OBS_STATE,
-    POLICY_POSTPROCESSOR_DEFAULT_NAME,
-    POLICY_PREPROCESSOR_DEFAULT_NAME,
-)
+from lerobot.utils.constants import OBS_STATE
 from lerobot.utils.import_utils import _transformers_available, require_package
 
 from .configuration_eo1 import EO1Config
@@ -242,14 +234,12 @@ def make_eo1_pre_post_processors(
 ]:
     """Build pre/post processor pipelines for EO1."""
 
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     input_steps: list[ProcessorStep] = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
+        steps.rename_observations,
+        steps.add_batch_dim,
+        steps.normalize,
         EO1ConversationTemplateStep(input_features=config.input_features, chunk_size=config.chunk_size),
         EO1QwenProcessorStep(
             processor_name=config.vlm_base,
@@ -257,27 +247,12 @@ def make_eo1_pre_post_processors(
             image_max_pixels=config.image_max_pixels,
             use_fast_processor=config.use_fast_processor,
         ),
-        DeviceProcessorStep(device=config.device),
+        steps.to_device,
     ]
 
     output_steps: list[ProcessorStep] = [
-        UnnormalizerProcessorStep(
-            features=config.output_features,
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-        DeviceProcessorStep(device="cpu"),
+        steps.unnormalize,
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/factory.py

@@ -17,6 +17,7 @@
 from __future__ import annotations
 
 import importlib
+import inspect
 import logging
 from typing import TYPE_CHECKING, Any, TypedDict, Unpack
 
@@ -44,26 +45,10 @@ from lerobot.utils.constants import (
 )
 from lerobot.utils.feature_utils import dataset_to_policy_features
 
-from .act.configuration_act import ACTConfig
-from .diffusion.configuration_diffusion import DiffusionConfig
-from .eo1.configuration_eo1 import EO1Config
 from .evo1.configuration_evo1 import Evo1Config
-from .fastwam.configuration_fastwam import FastWAMConfig
-from .gaussian_actor.configuration_gaussian_actor import GaussianActorConfig
 from .groot.configuration_groot import GrootConfig
-from .lingbot_va.configuration_lingbot_va import LingBotVAConfig
-from .molmoact2.configuration_molmoact2 import MolmoAct2Config
-from .multi_task_dit.configuration_multi_task_dit import MultiTaskDiTConfig
-from .pi0.configuration_pi0 import PI0Config
-from .pi05.configuration_pi05 import PI05Config
 from .pretrained import PreTrainedPolicy
-from .smolvla.configuration_smolvla import SmolVLAConfig
-from .tdmpc.configuration_tdmpc import TDMPCConfig
 from .utils import validate_visual_features_consistency
-from .vla_jepa.configuration_vla_jepa import VLAJEPAConfig
-from .vqbet.configuration_vqbet import VQBeTConfig
-from .wall_x.configuration_wall_x import WallXConfig
-from .xvla.configuration_xvla import XVLAConfig
 
 
 def _reconnect_relative_absolute_steps(
@@ -88,100 +73,23 @@ def get_policy_class(name: str) -> type[PreTrainedPolicy]:
     """
     Retrieves a policy class by its registered name.
 
-    This function uses dynamic imports to avoid loading all policy classes into memory
-    at once, improving startup time and reducing dependencies.
+    Resolution is convention-based: the draccus-registered config class of ``name`` is
+    looked up, its ``configuration_*`` module path is rewritten to ``modeling_*``, and
+    the ``<X>Policy`` class is imported from there. The modeling module is only imported
+    at call time, keeping heavy optional dependencies lazy. This works for both built-in
+    policies and third-party lerobot plugins (anything registered via
+    ``@PreTrainedConfig.register_subclass``).
 
     Args:
-        name: The name of the policy. Supported names are "tdmpc", "diffusion", "act",
-            "multi_task_dit", "vqbet", "pi0", "pi05", "gaussian_actor", "smolvla", "wall_x",
-            "molmoact2", "eo1", "evo1".
+        name: The registered name of the policy (e.g. "act", "diffusion", "pi0").
     Returns:
         The policy class corresponding to the given name.
 
     Raises:
-        NotImplementedError: If the policy name is not recognized.
+        ValueError: If the policy name is not registered.
+        ImportError: If the policy's optional dependencies are not installed.
     """
-    if name == "tdmpc":
-        from .tdmpc.modeling_tdmpc import TDMPCPolicy
-
-        return TDMPCPolicy
-    elif name == "diffusion":
-        from .diffusion.modeling_diffusion import DiffusionPolicy
-
-        return DiffusionPolicy
-    elif name == "act":
-        from .act.modeling_act import ACTPolicy
-
-        return ACTPolicy
-    elif name == "multi_task_dit":
-        from .multi_task_dit.modeling_multi_task_dit import MultiTaskDiTPolicy
-
-        return MultiTaskDiTPolicy
-    elif name == "vqbet":
-        from .vqbet.modeling_vqbet import VQBeTPolicy
-
-        return VQBeTPolicy
-    elif name == "pi0":
-        from .pi0.modeling_pi0 import PI0Policy
-
-        return PI0Policy
-    elif name == "pi0_fast":
-        from .pi0_fast.modeling_pi0_fast import PI0FastPolicy
-
-        return PI0FastPolicy
-    elif name == "pi05":
-        from .pi05.modeling_pi05 import PI05Policy
-
-        return PI05Policy
-    elif name == "gaussian_actor":
-        from .gaussian_actor.modeling_gaussian_actor import GaussianActorPolicy
-
-        return GaussianActorPolicy
-    elif name == "smolvla":
-        from .smolvla.modeling_smolvla import SmolVLAPolicy
-
-        return SmolVLAPolicy
-    elif name == "groot":
-        from .groot.modeling_groot import GrootPolicy
-
-        return GrootPolicy
-    elif name == "xvla":
-        from .xvla.modeling_xvla import XVLAPolicy
-
-        return XVLAPolicy
-    elif name == "wall_x":
-        from .wall_x.modeling_wall_x import WallXPolicy
-
-        return WallXPolicy
-    elif name == "eo1":
-        from .eo1.modeling_eo1 import EO1Policy
-
-        return EO1Policy
-    elif name == "molmoact2":
-        from .molmoact2.modeling_molmoact2 import MolmoAct2Policy
-
-        return MolmoAct2Policy
-    elif name == "vla_jepa":
-        from .vla_jepa.modeling_vla_jepa import VLAJEPAPolicy
-
-        return VLAJEPAPolicy
-    elif name == "lingbot_va":
-        from .lingbot_va.modeling_lingbot_va import LingBotVAPolicy
-
-        return LingBotVAPolicy
-    elif name == "fastwam":
-        from .fastwam.modeling_fastwam import FastWAMPolicy
-
-        return FastWAMPolicy
-    elif name == "evo1":
-        from .evo1.modeling_evo1 import Evo1Policy
-
-        return Evo1Policy
-    else:
-        try:
-            return _get_policy_cls_from_policy_name(name=name)
-        except Exception as e:
-            raise ValueError(f"Policy type '{name}' is not available.") from e
+    return _get_policy_cls_from_policy_name(name=name)
 
 
 def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
@@ -192,9 +100,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
     mapping a string identifier to the corresponding config class.
 
     Args:
-        policy_type: The type of the policy. Supported types include "tdmpc",
-                     "multi_task_dit", "diffusion", "act", "vqbet", "pi0", "pi05", "gaussian_actor",
-                     "smolvla", "wall_x", "molmoact2", "eo1", "evo1".
+        policy_type: The registered type of the policy (any name registered via
+                     ``@PreTrainedConfig.register_subclass``, e.g. "act", "diffusion", "pi0").
         **kwargs: Keyword arguments to be passed to the configuration class constructor.
 
     Returns:
@@ -203,48 +110,11 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
     Raises:
         ValueError: If the `policy_type` is not recognized.
     """
-    if policy_type == "tdmpc":
-        return TDMPCConfig(**kwargs)
-    elif policy_type == "diffusion":
-        return DiffusionConfig(**kwargs)
-    elif policy_type == "act":
-        return ACTConfig(**kwargs)
-    elif policy_type == "multi_task_dit":
-        return MultiTaskDiTConfig(**kwargs)
-    elif policy_type == "vqbet":
-        return VQBeTConfig(**kwargs)
-    elif policy_type == "pi0":
-        return PI0Config(**kwargs)
-    elif policy_type == "pi05":
-        return PI05Config(**kwargs)
-    elif policy_type == "gaussian_actor":
-        return GaussianActorConfig(**kwargs)
-    elif policy_type == "smolvla":
-        return SmolVLAConfig(**kwargs)
-    elif policy_type == "groot":
-        return GrootConfig(**kwargs)
-    elif policy_type == "xvla":
-        return XVLAConfig(**kwargs)
-    elif policy_type == "wall_x":
-        return WallXConfig(**kwargs)
-    elif policy_type == "eo1":
-        return EO1Config(**kwargs)
-    elif policy_type == "molmoact2":
-        return MolmoAct2Config(**kwargs)
-    elif policy_type == "vla_jepa":
-        return VLAJEPAConfig(**kwargs)
-    elif policy_type == "lingbot_va":
-        return LingBotVAConfig(**kwargs)
-    elif policy_type == "fastwam":
-        return FastWAMConfig(**kwargs)
-    elif policy_type == "evo1":
-        return Evo1Config(**kwargs)
-    else:
-        try:
-            config_cls = PreTrainedConfig.get_choice_class(policy_type)
-            return config_cls(**kwargs)
-        except Exception as e:
-            raise ValueError(f"Policy type '{policy_type}' is not available.") from e
+    try:
+        config_cls = PreTrainedConfig.get_choice_class(policy_type)
+    except Exception as e:
+        raise ValueError(f"Policy type '{policy_type}' is not available.") from e
+    return config_cls(**kwargs)
 
 
 class ProcessorConfigKwargs(TypedDict, total=False):
@@ -298,8 +168,7 @@ def make_pre_post_processors(
         A tuple containing the input (pre-processor) and output (post-processor) pipelines.
 
     Raises:
-        NotImplementedError: If a processor factory is not implemented for the given
-            policy configuration type.
+        ValueError: If no processor factory exists for the given policy configuration type.
     """
     if pretrained_path:
         if isinstance(policy_cfg, GrootConfig):
@@ -351,166 +220,13 @@ def make_pre_post_processors(
             )
         return preprocessor, postprocessor
 
-    # Create a new processor based on policy type
-    if isinstance(policy_cfg, TDMPCConfig):
-        from .tdmpc.processor_tdmpc import make_tdmpc_pre_post_processors
-
-        processors = make_tdmpc_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, DiffusionConfig):
-        from .diffusion.processor_diffusion import make_diffusion_pre_post_processors
-
-        processors = make_diffusion_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, ACTConfig):
-        from .act.processor_act import make_act_pre_post_processors
-
-        processors = make_act_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, MultiTaskDiTConfig):
-        from .multi_task_dit.processor_multi_task_dit import (
-            make_multi_task_dit_pre_post_processors,
-        )
-
-        processors = make_multi_task_dit_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, VQBeTConfig):
-        from .vqbet.processor_vqbet import make_vqbet_pre_post_processors
-
-        processors = make_vqbet_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, PI0Config):
-        from .pi0.processor_pi0 import make_pi0_pre_post_processors
-
-        processors = make_pi0_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, PI05Config):
-        from .pi05.processor_pi05 import make_pi05_pre_post_processors
-
-        processors = make_pi05_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, GaussianActorConfig):
-        from .gaussian_actor.processor_gaussian_actor import make_gaussian_actor_pre_post_processors
-
-        processors = make_gaussian_actor_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, SmolVLAConfig):
-        from .smolvla.processor_smolvla import make_smolvla_pre_post_processors
-
-        processors = make_smolvla_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, GrootConfig):
-        from .groot.processor_groot import make_groot_pre_post_processors
-
-        processors = make_groot_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-            dataset_meta=kwargs.get("dataset_meta"),
-        )
-
-    elif isinstance(policy_cfg, XVLAConfig):
-        from .xvla.processor_xvla import (
-            make_xvla_pre_post_processors,
-        )
-
-        processors = make_xvla_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, WallXConfig):
-        from .wall_x.processor_wall_x import make_wall_x_pre_post_processors
-
-        processors = make_wall_x_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, EO1Config):
-        from .eo1.processor_eo1 import make_eo1_pre_post_processors
-
-        processors = make_eo1_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-    elif isinstance(policy_cfg, Evo1Config):
-        from .evo1.processor_evo1 import make_evo1_pre_post_processors
-
-        processors = make_evo1_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, MolmoAct2Config):
-        from .molmoact2.processor_molmoact2 import make_molmoact2_pre_post_processors
-
-        processors = make_molmoact2_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-            dataset_meta=kwargs.get("dataset_meta"),
-        )
-
-    elif isinstance(policy_cfg, VLAJEPAConfig):
-        from .vla_jepa.processor_vla_jepa import make_vla_jepa_pre_post_processors
-
-        processors = make_vla_jepa_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, LingBotVAConfig):
-        from .lingbot_va.processor_lingbot_va import make_lingbot_va_pre_post_processors
-
-        processors = make_lingbot_va_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    elif isinstance(policy_cfg, FastWAMConfig):
-        from .fastwam.processor_fastwam import make_fastwam_pre_post_processors
-
-        processors = make_fastwam_pre_post_processors(
-            config=policy_cfg,
-            dataset_stats=kwargs.get("dataset_stats"),
-        )
-
-    else:
-        try:
-            processors = _make_processors_from_policy_config(
-                config=policy_cfg,
-                dataset_stats=kwargs.get("dataset_stats"),
-            )
-        except Exception as e:
-            raise ValueError(f"Processor for policy type '{policy_cfg.type}' is not implemented.") from e
-
-    return processors
+    # Create new processors from the policy config, resolving the per-policy factory
+    # function by naming convention (lazy import keeps optional dependencies optional).
+    return _make_processors_from_policy_config(
+        config=policy_cfg,
+        dataset_stats=kwargs.get("dataset_stats"),
+        dataset_meta=kwargs.get("dataset_meta"),
+    )
 
 
 def make_policy(
@@ -654,10 +370,12 @@ def make_policy(
     return policy
 
 
-def _get_policy_cls_from_policy_name(name: str) -> type[PreTrainedConfig]:
+def _get_policy_cls_from_policy_name(name: str) -> type[PreTrainedPolicy]:
     """Get policy class from its registered name using dynamic imports.
 
-    This is used as a helper function to import policies from 3rd party lerobot plugins.
+    Works for built-in policies and 3rd party lerobot plugins alike: the config class
+    registered under ``name`` is resolved via the draccus ChoiceRegistry, and the policy
+    class is imported from the sibling ``modeling_*`` module by naming convention.
 
     Args:
         name: The name of the policy.
@@ -683,22 +401,39 @@ def _get_policy_cls_from_policy_name(name: str) -> type[PreTrainedConfig]:
         "configuration_", "modeling_"
     )  # e.g., configuration_diffusion -> modeling_diffusion
 
-    module = importlib.import_module(module_path)
-    policy_cls = getattr(module, cls_name)
+    try:
+        module = importlib.import_module(module_path)
+    except ModuleNotFoundError as e:
+        if e.name == module_path:
+            # The modeling_* module itself does not exist for this policy type. A missing
+            # optional dependency inside an existing module propagates unchanged instead,
+            # so its actionable install hint stays visible.
+            raise ValueError(f"Policy class for '{name}' is not implemented.") from e
+        raise
+    policy_cls = getattr(module, cls_name, None)
+    if policy_cls is None:
+        raise ValueError(
+            f"Policy class '{cls_name}' not found in '{module_path}'. "
+            f"Policies must expose '<Name>Policy' in the sibling 'modeling_*' module by naming convention."
+        )
     return policy_cls
 
 
 def _make_processors_from_policy_config(
     config: PreTrainedConfig,
     dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    dataset_meta: Any | None = None,
 ) -> tuple[Any, Any]:
     """Create pre- and post-processors from a policy configuration using dynamic imports.
 
-    This is used as a helper function to import processor factories from 3rd party lerobot plugins.
+    Resolves ``make_{type}_pre_post_processors`` from the policy's ``processor_*`` module
+    by naming convention. Works for built-in policies and 3rd party lerobot plugins.
 
     Args:
         config: The policy configuration object.
         dataset_stats: Dataset statistics for normalization.
+        dataset_meta: Dataset metadata, forwarded only to factories that declare a
+            ``dataset_meta`` parameter (e.g. groot, molmoact2).
     Returns:
         A tuple containing the input (pre-processor) and output (post-processor) pipelines.
     """
@@ -711,6 +446,19 @@ def _make_processors_from_policy_config(
     logging.debug(
         f"Instantiating pre/post processors using function '{function_name}' from module '{module_path}'"
     )
-    module = importlib.import_module(module_path)
-    function = getattr(module, function_name)
-    return function(config, dataset_stats=dataset_stats)
+    try:
+        module = importlib.import_module(module_path)
+    except ModuleNotFoundError as e:
+        if e.name == module_path:
+            # The processor_* module itself does not exist for this policy type. A missing
+            # optional dependency inside an existing module propagates unchanged instead,
+            # so its actionable install hint stays visible.
+            raise ValueError(f"Processor for policy type '{policy_type}' is not implemented.") from e
+        raise
+    function = getattr(module, function_name, None)
+    if function is None:
+        raise ValueError(f"Processor for policy type '{policy_type}' is not implemented.")
+    call_kwargs: dict[str, Any] = {"dataset_stats": dataset_stats}
+    if "dataset_meta" in inspect.signature(function).parameters:
+        call_kwargs["dataset_meta"] = dataset_meta
+    return function(config, **call_kwargs)

src/lerobot/policies/fastwam/processor_fastwam.py

@@ -22,20 +22,11 @@ import torch
 from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
     ActionProcessorStep,
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStepRegistry,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
-)
-from lerobot.utils.constants import (
-    POLICY_POSTPROCESSOR_DEFAULT_NAME,
-    POLICY_PREPROCESSOR_DEFAULT_NAME,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
 
 from .configuration_fastwam import FastWAMConfig
@@ -105,38 +96,20 @@ def make_fastwam_pre_post_processors(
     # anyway) and unsafe across fine-tuning: its `resize_size` would be inherited from the base
     # checkpoint's camera geometry, not this dataset's, making the concatenation N_cameras x too wide.
 
+    steps = make_default_policy_processor_steps(config, normalization_stats, normalizer_device=config.device)
+
     input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=normalization_stats,
-            device=config.device,
-        ),
+        steps.rename_observations,
+        steps.add_batch_dim,
+        steps.to_device,
+        steps.normalize,
     ]
     output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features,
-            norm_map=config.normalization_mapping,
-            stats=normalization_stats,
-        ),
+        steps.unnormalize,
     ]
     if config.toggle_action_dimensions:
         output_steps.append(
             FastWAMActionToggleProcessorStep(toggle_dimensions=config.toggle_action_dimensions)
         )
-    output_steps.append(DeviceProcessorStep(device="cpu"))
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    output_steps.append(steps.to_cpu)
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/gaussian_actor/processor_gaussian_actor.py

@@ -20,17 +20,10 @@ from typing import Any
 import torch
 
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_pre_post_processors,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_gaussian_actor import GaussianActorConfig
 
@@ -62,33 +55,4 @@ def make_gaussian_actor_pre_post_processors(
     Returns:
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
-
-    # Add remaining processors
-    input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-    ]
-    output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
-    ]
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_default_pre_post_processors(config, dataset_stats)

src/lerobot/policies/lingbot_va/processor_lingbot_va.py

@@ -25,19 +25,12 @@ import torch
 
 from lerobot.configs.types import FeatureType, NormalizationMode
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStep,
-    RenameObservationsProcessorStep,
     UnnormalizerProcessorStep,
-)
-from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
-from lerobot.utils.constants import (
-    POLICY_POSTPROCESSOR_DEFAULT_NAME,
-    POLICY_PREPROCESSOR_DEFAULT_NAME,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
 
 from .configuration_lingbot_va import LingBotVAConfig
@@ -52,15 +45,13 @@ def make_lingbot_va_pre_post_processors(
 ]:
     """Build the pre/post processor pipelines for LingBot-VA."""
 
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     input_steps: list[ProcessorStep] = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-        DeviceProcessorStep(device=config.device),
+        steps.rename_observations,
+        steps.add_batch_dim,
+        steps.normalize,
+        steps.to_device,
     ]
 
     # Unnormalize actions from [-1, 1] to physical units (QUANTILES) using q01/q99 restored from the checkpoint.
@@ -70,18 +61,7 @@ def make_lingbot_va_pre_post_processors(
             norm_map={FeatureType.ACTION: NormalizationMode.QUANTILES},
             stats=dataset_stats,
         ),
-        DeviceProcessorStep(device="cpu"),
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/multi_task_dit/processor_multi_task_dit.py

@@ -19,18 +19,12 @@ from typing import Any
 import torch
 
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
     TokenizerProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_multi_task_dit import MultiTaskDiTConfig
 
@@ -66,9 +60,11 @@ def make_multi_task_dit_pre_post_processors(
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
 
+    steps = make_default_policy_processor_steps(config, dataset_stats, normalizer_device=config.device)
+
     input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
+        steps.rename_observations,
+        steps.add_batch_dim,
         TokenizerProcessorStep(
             tokenizer_name=config.text_encoder_name,
             padding=config.tokenizer_padding,
@@ -76,32 +72,12 @@ def make_multi_task_dit_pre_post_processors(
             max_length=config.tokenizer_max_length,
             truncation=config.tokenizer_truncation,
         ),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-            device=config.device,
-        ),
+        steps.to_device,
+        steps.normalize,
     ]
     output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features,
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-        DeviceProcessorStep(device="cpu"),
+        steps.unnormalize,
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/pi0/processor_pi0.py

@@ -21,22 +21,16 @@ import torch
 from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
     AbsoluteActionsProcessorStep,
-    AddBatchDimensionProcessorStep,
     ComplementaryDataProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStep,
     ProcessorStepRegistry,
     RelativeActionsProcessorStep,
-    RenameObservationsProcessorStep,
     TokenizerProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_pi0 import PI0Config
 
@@ -136,10 +130,12 @@ def make_pi0_pre_post_processors(
         action_names=getattr(config, "action_feature_names", None),
     )
 
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     # OpenPI order: raw → relative → normalize → model → unnormalize → absolute
     input_steps: list[ProcessorStep] = [
-        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
-        AddBatchDimensionProcessorStep(),
+        steps.rename_observations,  # To mimic the same processor as pretrained one
+        steps.add_batch_dim,
         Pi0NewLineProcessor(),  # Add newlines before tokenization for PaliGemma
         TokenizerProcessorStep(
             tokenizer_name="google/paligemma-3b-pt-224",
@@ -147,32 +143,15 @@ def make_pi0_pre_post_processors(
             padding_side="right",
             padding="max_length",
         ),
-        DeviceProcessorStep(device=config.device),
+        steps.to_device,
         relative_step,
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
+        steps.normalize,
     ]
 
     output_steps: list[ProcessorStep] = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
+        steps.unnormalize,
         AbsoluteActionsProcessorStep(enabled=config.use_relative_actions, relative_step=relative_step),
-        DeviceProcessorStep(device="cpu"),
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/pi05/processor_pi05.py

@@ -24,26 +24,17 @@ import torch
 from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
     AbsoluteActionsProcessorStep,
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStep,
     ProcessorStepRegistry,
     RelativeActionsProcessorStep,
-    RenameObservationsProcessorStep,
     TokenizerProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
 from lerobot.types import EnvTransition, TransitionKey
-from lerobot.utils.constants import (
-    OBS_STATE,
-    POLICY_POSTPROCESSOR_DEFAULT_NAME,
-    POLICY_PREPROCESSOR_DEFAULT_NAME,
-)
+from lerobot.utils.constants import OBS_STATE
 
 from .configuration_pi05 import PI05Config
 
@@ -135,18 +126,16 @@ def make_pi05_pre_post_processors(
         action_names=getattr(config, "action_feature_names", None),
     )
 
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     # OpenPI order: raw → relative → normalize → model → unnormalize → absolute
     input_steps: list[ProcessorStep] = [
-        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
-        AddBatchDimensionProcessorStep(),
+        steps.rename_observations,  # To mimic the same processor as pretrained one
+        steps.add_batch_dim,
         relative_step,
         # NOTE: NormalizerProcessorStep MUST come before Pi05PrepareStateTokenizerProcessorStep
         # because the tokenizer step expects normalized state in [-1, 1] range for discretization
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
+        steps.normalize,
         Pi05PrepareStateTokenizerProcessorStep(max_state_dim=config.max_state_dim),
         TokenizerProcessorStep(
             tokenizer_name="google/paligemma-3b-pt-224",
@@ -154,26 +143,13 @@ def make_pi05_pre_post_processors(
             padding_side="right",
             padding="max_length",
         ),
-        DeviceProcessorStep(device=config.device),
+        steps.to_device,
     ]
 
     output_steps: list[ProcessorStep] = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
+        steps.unnormalize,
         AbsoluteActionsProcessorStep(enabled=config.use_relative_actions, relative_step=relative_step),
-        DeviceProcessorStep(device="cpu"),
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/pi0_fast/processor_pi0_fast.py

@@ -25,26 +25,17 @@ from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
     AbsoluteActionsProcessorStep,
     ActionTokenizerProcessorStep,
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStep,
     ProcessorStepRegistry,
     RelativeActionsProcessorStep,
-    RenameObservationsProcessorStep,
     TokenizerProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
 from lerobot.types import EnvTransition, TransitionKey
-from lerobot.utils.constants import (
-    OBS_STATE,
-    POLICY_POSTPROCESSOR_DEFAULT_NAME,
-    POLICY_PREPROCESSOR_DEFAULT_NAME,
-)
+from lerobot.utils.constants import OBS_STATE
 
 from .configuration_pi0_fast import PI0FastConfig
 
@@ -135,6 +126,8 @@ def make_pi0_fast_pre_post_processors(
         action_names=getattr(config, "action_feature_names", None),
     )
 
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     # Pi0Fast order: relative → normalize → tokenize → model → unnormalize → absolute
     # This matches pi0/pi0.5: RelativeActionsProcessorStep runs first on raw absolute actions,
     # caching the raw state. NormalizerProcessorStep then normalizes the raw relative actions,
@@ -144,14 +137,10 @@ def make_pi0_fast_pre_post_processors(
     # before Pi0FastPrepareStateAndLanguageTokenizerProcessorStep, so the state tokenizer
     # continues to receive normalized state in [-1, 1] as expected.
     input_steps: list[ProcessorStep] = [
-        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
-        AddBatchDimensionProcessorStep(),
+        steps.rename_observations,  # To mimic the same processor as pretrained one
+        steps.add_batch_dim,
         relative_step,
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
+        steps.normalize,
         Pi0FastPrepareStateAndLanguageTokenizerProcessorStep(max_state_dim=config.max_state_dim),
         TokenizerProcessorStep(
             tokenizer_name=config.text_tokenizer_name,
@@ -165,26 +154,13 @@ def make_pi0_fast_pre_post_processors(
             fast_skip_tokens=config.fast_skip_tokens,
             paligemma_tokenizer_name=config.text_tokenizer_name,
         ),
-        DeviceProcessorStep(device=config.device),
+        steps.to_device,
     ]
 
     output_steps: list[ProcessorStep] = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
+        steps.unnormalize,
         AbsoluteActionsProcessorStep(enabled=config.use_relative_actions, relative_step=relative_step),
-        DeviceProcessorStep(device="cpu"),
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/smolvla/processor_smolvla.py

@@ -19,19 +19,13 @@ from typing import Any
 import torch
 
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
     NewLineTaskProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
     TokenizerProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_smolvla import SmolVLAConfig
 
@@ -66,9 +60,11 @@ def make_smolvla_pre_post_processors(
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
 
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),  # To mimic the same processor as pretrained one
-        AddBatchDimensionProcessorStep(),
+        steps.rename_observations,  # To mimic the same processor as pretrained one
+        steps.add_batch_dim,
         NewLineTaskProcessorStep(),
         TokenizerProcessorStep(
             tokenizer_name=config.vlm_model_name,
@@ -76,28 +72,11 @@ def make_smolvla_pre_post_processors(
             padding_side="right",
             max_length=config.tokenizer_max_length,
         ),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
+        steps.to_device,
+        steps.normalize,
     ]
     output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
+        steps.unnormalize,
+        steps.to_cpu,
     ]
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/tdmpc/processor_tdmpc.py

@@ -19,17 +19,10 @@ from typing import Any
 import torch
 
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_pre_post_processors,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_tdmpc import TDMPCConfig
 
@@ -61,32 +54,4 @@ def make_tdmpc_pre_post_processors(
     Returns:
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
-
-    input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-    ]
-    output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
-    ]
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_default_pre_post_processors(config, dataset_stats)

src/lerobot/policies/vla_jepa/processor_vla_jepa.py

@@ -20,20 +20,16 @@ import torch
 
 from lerobot.policies.vla_jepa.configuration_vla_jepa import VLAJEPAConfig
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
     EnvTransition,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStep,
     ProcessorStepRegistry,
-    RenameObservationsProcessorStep,
     TransitionKey,
     UnnormalizerProcessorStep,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
-from lerobot.processor.converters import policy_action_to_transition, transition_to_policy_action
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 
 @ProcessorStepRegistry.register(name="vla_jepa_clip_actions")
@@ -112,15 +108,12 @@ def make_vla_jepa_pre_post_processors(
     PolicyProcessorPipeline[PolicyAction, PolicyAction],
 ]:
     features = {**config.input_features, **config.output_features}
+    steps = make_default_policy_processor_steps(config, dataset_stats)
     input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features=features,
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
+        steps.rename_observations,
+        steps.add_batch_dim,
+        steps.to_device,
+        steps.normalize,
     ]
     output_steps: list[ProcessorStep] = []
     if config.clip_normalized_actions:
@@ -129,6 +122,8 @@ def make_vla_jepa_pre_post_processors(
         output_steps.append(
             PreSnapGripperProcessorStep(gripper_dim=config.gripper_dim, threshold=config.gripper_threshold)
         )
+    # NOTE: unlike the default policy unnormalizer (output features only), VLA-JEPA
+    # unnormalizes over BOTH input and output features.
     output_steps.append(
         UnnormalizerProcessorStep(
             features=features,
@@ -140,16 +135,5 @@ def make_vla_jepa_pre_post_processors(
         output_steps.append(
             BinarizeGripperProcessorStep(gripper_dim=config.gripper_dim, threshold=config.gripper_threshold)
         )
-    output_steps.append(DeviceProcessorStep(device="cpu"))
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    output_steps.append(steps.to_cpu)
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)

src/lerobot/policies/vqbet/processor_vqbet.py

@@ -20,17 +20,10 @@ from typing import Any
 import torch
 
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_pre_post_processors,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_vqbet import VQBeTConfig
 
@@ -62,32 +55,4 @@ def make_vqbet_pre_post_processors(
     Returns:
         A tuple containing the configured pre-processor and post-processor pipelines.
     """
-
-    input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),  # Let the possibility to the user to rename the keys
-        AddBatchDimensionProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-    ]
-    output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
-    ]
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_default_pre_post_processors(config, dataset_stats)

src/lerobot/policies/wall_x/configuration_wall_x.py

@@ -58,10 +58,14 @@ class WallXConfig(PreTrainedConfig):
     # Action prediction mode: "diffusion" or "fast"
     prediction_mode: str = "diffusion"
 
-    # Attention Implementation, options: "eager", "flash_attention_2", "sdpa"
-    # NOTE: flash-attn==2.7.4.post1 is required for flash_attention_2 implementation
+    # Wall-X's bidirectional action-token islands currently require eager attention.
     attn_implementation: str = "eager"
 
+    # Vision attention is independent from the text action-token mask. ``auto`` uses
+    # PyTorch's packed variable-length attention when the runtime supports it and
+    # otherwise falls back to the native per-chunk SDPA implementation.
+    vision_attn_implementation: str = "auto"
+
     # ==================== Optimizer Presets ====================
     optimizer_lr: float = 2e-5
     optimizer_betas: tuple[float, float] = (0.9, 0.95)
@@ -86,6 +90,18 @@ class WallXConfig(PreTrainedConfig):
         if self.prediction_mode not in ["diffusion", "fast"]:
             raise ValueError(f"prediction_mode must be 'diffusion' or 'fast', got {self.prediction_mode}")
 
+        if self.attn_implementation != "eager":
+            raise ValueError(
+                "Wall-X currently supports only attn_implementation='eager' because its "
+                "bidirectional action-token islands require an explicit attention mask."
+            )
+
+        if self.vision_attn_implementation not in {"auto", "sdpa", "varlen"}:
+            raise ValueError(
+                "vision_attn_implementation must be one of 'auto', 'sdpa', or 'varlen', got "
+                f"{self.vision_attn_implementation!r}"
+            )
+
         # Assign use_fast_tokenizer based on prediction_mode
         if self.prediction_mode == "fast":
             self.use_fast_tokenizer = True

src/lerobot/policies/wall_x/modeling_wall_x.py

@@ -43,11 +43,14 @@ from typing import TYPE_CHECKING, Any
 import numpy as np
 import torch
 import torch.nn as nn
-import torch.nn.functional as F
-from PIL import Image
+import torch.nn.functional as functional
+from safetensors import SafetensorError
+from safetensors.torch import load_file
 from torch import Tensor
 from torch.distributions import Beta
 from torch.nn import CrossEntropyLoss
+from torchvision.transforms import InterpolationMode
+from torchvision.transforms.v2 import functional as tv_functional
 
 from lerobot.utils.constants import ACTION, OBS_STATE
 from lerobot.utils.import_utils import (
@@ -74,17 +77,17 @@ if TYPE_CHECKING or _wallx_deps_available:
     from qwen_vl_utils.vision_process import smart_resize
     from torchdiffeq import odeint
     from transformers import AutoProcessor, BatchFeature
-    from transformers.cache_utils import StaticCache
     from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
+        Qwen2_5_VisionTransformerPretrainedModel,
         Qwen2_5_VLForConditionalGeneration,
     )
-    from transformers.utils import is_torchdynamo_compiling
+    from transformers.utils import cached_file, is_torchdynamo_compiling
 
-    from .qwen_model.configuration_qwen2_5_vl import Qwen2_5_VLConfig
-    from .qwen_model.qwen2_5_vl_moe import (
-        Qwen2_5_VisionTransformerPretrainedModel,
+    from .qwen_model import (
         Qwen2_5_VLACausalLMOutputWithPast,
+        Qwen2_5_VLConfig,
         Qwen2_5_VLMoEModel,
+        configure_wall_x_vision_attention,
     )
 else:
     LoraConfig = None
@@ -93,13 +96,14 @@ else:
     odeint = None
     AutoProcessor = None
     BatchFeature = None
-    StaticCache = None
     Qwen2_5_VLForConditionalGeneration = None
+    cached_file = None
     is_torchdynamo_compiling = None
     Qwen2_5_VLConfig = None
     Qwen2_5_VisionTransformerPretrainedModel = None
     Qwen2_5_VLACausalLMOutputWithPast = None
     Qwen2_5_VLMoEModel = None
+    configure_wall_x_vision_attention = None
 
 from .utils import (
     get_wallx_normal_text,
@@ -111,6 +115,75 @@ from .utils import (
 logger = logging.getLogger(__name__)
 
 
+def _wall_x_resize_dimensions(height: int, width: int) -> tuple[int, int, int, int]:
+    """Return the intermediate and final Wall-X resize dimensions as ``(H, W, H, W)``."""
+    if RESOLUTION == -1:
+        intermediate_height, intermediate_width = height, width
+    elif width > height:
+        intermediate_width = RESOLUTION
+        intermediate_height = int(RESOLUTION * height / width)
+    else:
+        intermediate_height = RESOLUTION
+        intermediate_width = int(RESOLUTION * width / height)
+
+    resized_height, resized_width = smart_resize(
+        intermediate_height,
+        intermediate_width,
+        factor=IMAGE_FACTOR,
+        min_pixels=MIN_PIXELS,
+        max_pixels=MAX_PIXELS,
+    )
+    return intermediate_height, intermediate_width, resized_height, resized_width
+
+
+def _resize_wall_x_image_batch(images: Tensor) -> tuple[Tensor, tuple[int, int, int, int]]:
+    """Quantize and resize a BCHW camera batch without leaving its current device."""
+    if images.ndim != 4:
+        raise ValueError(f"Wall-X images must be BCHW tensors, got shape {tuple(images.shape)}")
+
+    original_height, original_width = images.shape[-2:]
+    intermediate_height, intermediate_width, resized_height, resized_width = _wall_x_resize_dimensions(
+        original_height, original_width
+    )
+
+    if images.is_floating_point():
+        # Match the previous PIL path, which quantized via `(image * 255).to(torch.uint8)`.
+        images = (images * 255).to(torch.uint8)
+    elif images.dtype != torch.uint8:
+        raise TypeError(f"Wall-X images must be floating point or uint8, got {images.dtype}")
+
+    if images.shape[-2:] != (intermediate_height, intermediate_width):
+        images = tv_functional.resize(
+            images,
+            [intermediate_height, intermediate_width],
+            interpolation=InterpolationMode.BICUBIC,
+            antialias=True,
+        )
+    if images.shape[-2:] != (resized_height, resized_width):
+        images = tv_functional.resize(
+            images,
+            [resized_height, resized_width],
+            interpolation=InterpolationMode.BICUBIC,
+            antialias=True,
+        )
+
+    return images, (original_height, original_width, resized_height, resized_width)
+
+
+def _prepare_wall_x_image_inputs(
+    batch: dict[str, Any], img_keys: list[str]
+) -> tuple[list[list[Tensor]], dict[str, tuple[int, int, int, int]]]:
+    """Resize each camera as a batch, then restore sample-major/camera-minor ordering."""
+    resized_by_key: dict[str, Tensor] = {}
+    dimensions_by_key: dict[str, tuple[int, int, int, int]] = {}
+    for key in img_keys:
+        resized_by_key[key], dimensions_by_key[key] = _resize_wall_x_image_batch(batch[key])
+
+    batch_size = batch[img_keys[0]].shape[0]
+    image_inputs = [[resized_by_key[key][i] for key in img_keys] for i in range(batch_size)]
+    return image_inputs, dimensions_by_key
+
+
 class SinusoidalPosEmb(nn.Module):
     """Sinusoidal positional embedding for diffusion timesteps."""
 
@@ -246,7 +319,7 @@ class ActionHead(nn.Module):
         flow = flow.to(torch.float32)
 
         action_pred = self.action_proj_back(action_hidden_states)
-        loss = F.mse_loss(action_pred, flow, reduction="none")
+        loss = functional.mse_loss(action_pred, flow, reduction="none")
 
         if dof_mask is not None:
             dof_mask = dof_mask.reshape(-1, dof_mask.shape[-1]).to(torch.float32)
@@ -254,7 +327,7 @@ class ActionHead(nn.Module):
 
         return loss
 
-    def proprioception_proj(self, proprioception, dof_mask=None, use_history=False):
+    def proprioception_proj(self, proprioception, dof_mask=None):
         """Project proprioceptive data to hidden space."""
         # Ensure proper device and dtype alignment
         proprioception = proprioception.to(device=self.propri_proj.weight.device).to(
@@ -264,10 +337,7 @@ class ActionHead(nn.Module):
         if dof_mask is not None:
             # Concatenate proprioception with DOF mask
             # TODO: Use variable-based dimension checking for better flexibility
-            if use_history:
-                proprioception = torch.cat([proprioception, dof_mask], dim=-1)
-            else:
-                proprioception = torch.cat([proprioception, dof_mask], dim=-1)
+            proprioception = torch.cat([proprioception, dof_mask], dim=-1)
 
         proprioception = proprioception.to(device=self.propri_proj.weight.device).to(
             dtype=self.propri_proj.weight.dtype
@@ -281,7 +351,7 @@ class ActionHead(nn.Module):
 _Qwen2_5_VLForAction_Base = Qwen2_5_VLForConditionalGeneration if _wallx_deps_available else nn.Module
 
 
-class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
+class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):  # noqa: N801
     """
     Qwen2.5 Vision-Language Mixture of Experts model for action processing.
 
@@ -305,6 +375,7 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
         config=None,
         action_tokenizer_path=None,
         attn_implementation: str = "eager",
+        vision_attn_implementation: str = "auto",
         cache_dir: str | PathLike | None = None,
         force_download: bool = False,
         local_files_only: bool = False,
@@ -321,11 +392,14 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
             config_path (str, optional): Configuration file path, if None will look for qwen25_config.json in pretrained_model_path
             action_tokenizer_path (str, optional): Action tokenizer path, if None will load from default config
             attn_implementation (str, optional): Attention implementation, if None will load from default config
+            vision_attn_implementation (str, optional): Vision attention backend. ``auto`` uses packed
+                variable-length attention when supported and otherwise falls back to SDPA.
             **kwargs: Additional arguments
 
         Returns:
             Qwen2_5_VLMoEForAction: Loaded model instance
         """
+        Qwen2_5_VLMoEModel._require_eager_attention(attn_implementation)
         if config is None:
             config = cls.config_class.from_pretrained(
                 pretrained_name_or_path,
@@ -339,7 +413,15 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
             )
         if attn_implementation is not None:
             config._attn_implementation = attn_implementation
-        processor = AutoProcessor.from_pretrained(pretrained_name_or_path, use_fast=True)
+        processor = AutoProcessor.from_pretrained(
+            pretrained_name_or_path,
+            cache_dir=cache_dir,
+            force_download=force_download,
+            local_files_only=local_files_only,
+            token=token,
+            revision=revision,
+            use_fast=True,
+        )
         if action_tokenizer_path is not None:
             action_tokenizer = AutoProcessor.from_pretrained(action_tokenizer_path, trust_remote_code=True)
             processor.action_processor = action_tokenizer
@@ -351,41 +433,41 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
         config.text_config.pad_token_id = processor.tokenizer.pad_token_id
 
         # Initialize model with configuration and processor
-        model = cls(config, processor=processor, action_tokenizer=action_tokenizer, **kwargs)
+        model = cls(
+            config,
+            processor=processor,
+            action_tokenizer=action_tokenizer,
+            vision_attn_implementation=vision_attn_implementation,
+            **kwargs,
+        )
 
         # Resize token embeddings to match processor tokenizer vocabulary size
         model.resize_token_embeddings(len(processor.tokenizer))
 
-        # Try to load the model.safetensors file
-        print(f"Loading model from: {pretrained_name_or_path}")
+        logger.info("Loading Wall-X model from %s", pretrained_name_or_path)
         try:
-            from transformers.utils import cached_file
-
-            # Try safetensors first
             resolved_file = cached_file(
                 pretrained_name_or_path,
                 "model.safetensors",
-                cache_dir=kwargs.get("cache_dir"),
-                force_download=kwargs.get("force_download", False),
+                cache_dir=cache_dir,
+                force_download=force_download,
                 resume_download=kwargs.get("resume_download"),
                 proxies=kwargs.get("proxies"),
-                token=kwargs.get("token"),
-                revision=kwargs.get("revision"),
-                local_files_only=kwargs.get("local_files_only", False),
+                token=token,
+                revision=revision,
+                local_files_only=local_files_only,
             )
-            from safetensors.torch import load_file
-
             sd = load_file(resolved_file)
-            print("✓ Loaded state dict from model.safetensors")
-        except Exception as e:
-            print(f"Could not load state dict from remote files: {e}")
-            print("Returning model without loading pretrained weights")
-            return model
+        except (OSError, SafetensorError) as error:
+            raise OSError(
+                f"Failed to load pretrained Wall-X weights from {pretrained_name_or_path!r}"
+            ) from error
+        logger.info("Loaded Wall-X state dict from model.safetensors")
 
         state_dict = {}
         # filter normalizer statistic params
         del_keys = []
-        for key in sd.keys():
+        for key in sd:
             if "action_preprocessor.normalizer" in key:
                 del_keys.append(key)
         for key in del_keys:
@@ -404,6 +486,7 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
         action_tokenizer=None,
         action_mapper=None,
         flow_loss_weight=1.0,
+        vision_attn_implementation: str = "auto",
     ):
         """
         Initialize the Qwen2.5 VLMoE model for action processing.
@@ -416,10 +499,16 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
             action_mapper: Action mapping utility
             flow_loss_weight (float): Weight for flow loss computation
         """
+        Qwen2_5_VLMoEModel._require_eager_attention(config._attn_implementation)
+        config._attn_implementation = "eager"
+        # Text needs eager attention for action-token islands. Vision has no such
+        # constraint, so keep its portable native fallback on SDPA.
+        config.vision_config._attn_implementation = "sdpa"
         super().__init__(config)
 
         # Initialize vision transformer and language model components
         self.visual = Qwen2_5_VisionTransformerPretrainedModel._from_config(config.vision_config)
+        configure_wall_x_vision_attention(self.visual, vision_attn_implementation)
         self.model = Qwen2_5_VLMoEModel(config)
         self.vocab_size = config.vocab_size
         self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
@@ -457,7 +546,7 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
 
         params_to_keep_float32 = []
 
-        for name, param in self.named_parameters():
+        for name, _param in self.named_parameters():
             if "input_layernorm" in name or "post_attention_layernorm" in name or "model.norm" in name:
                 params_to_keep_float32.append(name)
             if "action_preprocessor" in name:
@@ -491,7 +580,7 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
             "action_token_id": action_token_id,
         }
 
-    def add_lora(self, r=8, lora_alpha=32, target_modules=["q_proj", "v_proj"], lora_dropout=0.1):
+    def add_lora(self, r=8, lora_alpha=32, target_modules=None, lora_dropout=0.1):
         """
         Add LoRA (Low-Rank Adaptation) adapters to the model.
 
@@ -501,6 +590,9 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
             target_modules (list): List of module names to apply LoRA to
             lora_dropout (float): Dropout probability for LoRA layers
         """
+        if target_modules is None:
+            target_modules = ["q_proj", "v_proj"]
+
         config = LoraConfig(
             r=r,
             lora_alpha=lora_alpha,
@@ -795,6 +887,9 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
         )
         return_dict = return_dict if return_dict is not None else self.config.use_return_dict
 
+        if rope_deltas is not None:
+            self.rope_deltas = rope_deltas
+
         # Calculate RoPE position IDs if not provided
         # Note: Cannot calculate rope deltas with 4D attention mask. TODO: Fix this limitation
         if position_ids is None and (attention_mask is None or attention_mask.ndim == 2):
@@ -833,7 +928,7 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
             # Process image embeddings
             if pixel_values is not None:
                 pixel_values = pixel_values.type(self.visual.dtype)
-                image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
+                image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw).pooler_output
                 mask = input_ids == self.config.image_token_id
                 mask_unsqueezed = mask.unsqueeze(-1)
                 mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
@@ -845,7 +940,7 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
             # Process video embeddings
             if pixel_values_videos is not None:
                 pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
-                video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
+                video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw).pooler_output
                 n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
                 n_video_features = video_embeds.shape[0]
 
@@ -869,7 +964,6 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
                 proprioception = self.action_preprocessor.proprioception_proj(
                     proprioception,
                     agent_pos_mask,
-                    use_history=proprioception.shape[1] > 1,
                 )
                 mask = input_ids == self.action_token_id_set["propri_token_id"]
                 mask_unsqueezed = mask.unsqueeze(-1)
@@ -919,6 +1013,7 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
             output_attentions=output_attentions,
             output_hidden_states=output_hidden_states,
             return_dict=return_dict,
+            cache_position=cache_position,
         )
 
         hidden_states = outputs[0]
@@ -1107,7 +1202,7 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
             # Process image embeddings
             if pixel_values is not None:
                 pixel_values = pixel_values.type(self.visual.dtype)
-                image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
+                image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw).pooler_output
                 n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
                 n_image_features = image_embeds.shape[0]
 
@@ -1128,7 +1223,7 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
             # Process video embeddings
             if pixel_values_videos is not None:
                 pixel_values_videos = pixel_values_videos.type(self.visual.dtype)
-                video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
+                video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw).pooler_output
                 n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
                 n_video_features = video_embeds.shape[0]
 
@@ -1153,7 +1248,6 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
                 proprio_embed = self.action_preprocessor.proprioception_proj(
                     proprioception,
                     agent_pos_mask,
-                    use_history=proprioception.shape[1] > 1,
                 )
                 proprioception_mask = input_ids == self.action_token_id_set["propri_token_id"]
                 proprio_embed = proprio_embed.to(torch.bfloat16)
@@ -1202,25 +1296,37 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
 
         # Split input sequence for text and fast modes (not needed for diffusion)
         if predict_mode == "text" or predict_mode == "fast":
-            # Look for generation prompt tokens: <|im_start|>assistant
+            generation_prompt = "<|im_start|>assistant\n"
             generation_prompt_ids = torch.tensor(
-                [151644, 77091], device=input_ids.device, dtype=input_ids.dtype
-            )
-            matches = (input_ids[0, :-1] == generation_prompt_ids[0]) & (
-                input_ids[0, 1:] == generation_prompt_ids[1]
+                self.processor.tokenizer.encode(generation_prompt, add_special_tokens=False),
+                device=input_ids.device,
+                dtype=input_ids.dtype,
             )
+            prompt_length = generation_prompt_ids.numel()
+            if prompt_length == 0:
+                raise ValueError(f"Tokenizer produced no tokens for generation prompt {generation_prompt!r}")
+            if input_ids.shape[1] < prompt_length:
+                matches = torch.empty(0, device=input_ids.device, dtype=torch.bool)
+            else:
+                matches = (
+                    input_ids[0]
+                    .unfold(dimension=0, size=prompt_length, step=1)
+                    .eq(generation_prompt_ids)
+                    .all(dim=-1)
+                )
 
             if matches.any():
                 split_pos = torch.nonzero(matches, as_tuple=True)[0][0].item()
+                prompt_end = split_pos + prompt_length
                 # Extract ground truth output tokens (including newline)
-                gt_output_ids = input_ids[:, split_pos + 3 :]
+                gt_output_ids = input_ids[:, prompt_end:]
                 # Remove output part from input, keeping prompt
-                input_ids = input_ids[:, : split_pos + 3]
-                inputs_embeds = inputs_embeds[:, : split_pos + 3, :]
+                input_ids = input_ids[:, :prompt_end]
+                inputs_embeds = inputs_embeds[:, :prompt_end, :]
                 if attention_mask is not None:
-                    attention_mask = attention_mask[:, : split_pos + 3]
+                    attention_mask = attention_mask[:, :prompt_end]
                 if labels is not None:
-                    labels = labels[:, split_pos + 3 :]
+                    labels = labels[:, prompt_end:]
             else:
                 raise ValueError(
                     "input_ids does not contain the generation prompt tokens <|im_start|>assistant"
@@ -1255,7 +1361,7 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
                 use_cache=True,
                 pad_token_id=self.processor.tokenizer.pad_token_id,
                 temperature=(1.0 if not re_generate else 0.7),  # Higher temperature for regeneration
-                do_sample=(False if not re_generate else True),  # Enable sampling for regeneration
+                do_sample=re_generate,  # Enable sampling for regeneration
             )
 
             # Decode generated and ground truth text
@@ -1524,27 +1630,6 @@ class Qwen2_5_VLMoEForAction(_Qwen2_5_VLForAction_Base):
         else:
             model_inputs = {"input_ids": input_ids, "inputs_embeds": None}
 
-        # Prepare 4D causal attention mask for static cache
-        if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
-            if model_inputs["inputs_embeds"] is not None:
-                batch_size, sequence_length, _ = inputs_embeds.shape
-                device = inputs_embeds.device
-            else:
-                batch_size, sequence_length = input_ids.shape
-                device = input_ids.device
-
-            attention_mask = self.model._prepare_4d_causal_attention_mask_with_cache_position(
-                attention_mask,
-                sequence_length=sequence_length,
-                target_length=past_key_values.get_max_cache_shape(),
-                dtype=self.lm_head.weight.dtype,
-                device=device,
-                cache_position=cache_position,
-                batch_size=batch_size,
-                config=self.config,
-                past_key_values=past_key_values,
-            )
-
         # Assemble all model inputs for generation
         model_inputs.update(
             {
@@ -1749,6 +1834,7 @@ class WallXPolicy(PreTrainedPolicy):
             pretrained_name_or_path=config.pretrained_name_or_path,
             action_tokenizer_path=config.action_tokenizer_path,
             attn_implementation=config.attn_implementation,
+            vision_attn_implementation=config.vision_attn_implementation,
         )
         self.model.to(config.device)
         self.model.to_bfloat16_for_selected_params()
@@ -1768,6 +1854,8 @@ class WallXPolicy(PreTrainedPolicy):
     def preprocess_inputs(
         self,
         batch: dict[str, Any],
+        *,
+        compute_position_ids: bool = False,
     ) -> BatchFeature:
         """
         Convert a batch of LeRobot dataset items to Wall-X model input format.
@@ -1789,50 +1877,21 @@ class WallXPolicy(PreTrainedPolicy):
         # Get batch size from state tensor
         batch_size = batch[OBS_STATE].shape[0]
 
-        # ==================== PROCESS ALL SAMPLES ====================
-        all_image_inputs = []
-        all_texts = []
-
         # Find image keys in batch
         img_keys = [key for key in self.config.image_features if key in batch]
+        if not img_keys:
+            raise ValueError("Wall-X requires at least one image feature in each batch")
+
+        # Resize one camera batch at a time on the tensors' current device. Reassembling
+        # sample-major keeps image_grid_thw aligned with each sample's image placeholders.
+        all_image_inputs, dimensions_by_key = _prepare_wall_x_image_inputs(batch, img_keys)
+        all_texts = []
+
+        # Preserve the existing grounding behavior for multi-camera inputs: the old camera
+        # loop left these values set to the final configured camera's dimensions.
+        orig_height, orig_width, resized_height, resized_width = dimensions_by_key[img_keys[-1]]
 
         for i in range(batch_size):
-            # Vision preprocessing per sample
-            processed_frames = []
-            orig_height, orig_width = None, None
-            resized_height, resized_width = None, None
-
-            for key in img_keys:
-                current_obs = batch[key][i].clone()  # (C, H, W)
-                if current_obs.dim() == 3:
-                    current_obs = current_obs.permute(1, 2, 0)  # (H, W, C)
-
-                img_pil = Image.fromarray((current_obs * 255).to(torch.uint8).cpu().numpy())
-                orig_width, orig_height = img_pil.size
-
-                target_size = RESOLUTION
-                if target_size != -1:
-                    if orig_width > orig_height:
-                        new_width = target_size
-                        new_height = int(target_size * orig_height / orig_width)
-                    else:
-                        new_height = target_size
-                        new_width = int(target_size * orig_width / orig_height)
-                    img_pil = img_pil.resize((new_width, new_height))
-
-                current_width, current_height = img_pil.size
-                resized_height, resized_width = smart_resize(
-                    current_height,
-                    current_width,
-                    factor=IMAGE_FACTOR,
-                    min_pixels=MIN_PIXELS,
-                    max_pixels=MAX_PIXELS,
-                )
-                resized_img = img_pil.resize((resized_width, resized_height))
-                processed_frames.append(resized_img)
-
-            all_image_inputs.append(processed_frames)
-
             # Text preprocessing
             task_text = batch["task"][i] if isinstance(batch["task"], list) else batch["task"]
             instruction_info = {"instruction": task_text}
@@ -1859,8 +1918,8 @@ class WallXPolicy(PreTrainedPolicy):
         agent_pos_mask = (~torch.isnan(agent_pos)).float()
         agent_pos = agent_pos.nan_to_num(nan=0.0)
 
-        if agent_pos.shape[-1] != 20:
-            pad_size = 20 - agent_pos.shape[-1]
+        if agent_pos.shape[-1] < self.config.max_state_dim:
+            pad_size = self.config.max_state_dim - agent_pos.shape[-1]
             agent_pos = torch.cat(
                 [
                     agent_pos,
@@ -1880,6 +1939,10 @@ class WallXPolicy(PreTrainedPolicy):
                 ],
                 dim=-1,
             )
+        elif agent_pos.shape[-1] > self.config.max_state_dim:
+            raise ValueError(
+                f"State dimension {agent_pos.shape[-1]} exceeds max_state_dim {self.config.max_state_dim}"
+            )
 
         # ==================== PROCESS ACTIONS ====================
         action = batch.get(ACTION)  # (batch_size, chunk_size, action_dim)
@@ -1889,8 +1952,8 @@ class WallXPolicy(PreTrainedPolicy):
             dof_mask = (~torch.isnan(action)).float()
             action = action.nan_to_num(nan=0.0)
 
-            if action.shape[-1] != 20:
-                pad_size = 20 - action.shape[-1]
+            if action.shape[-1] < self.config.max_action_dim:
+                pad_size = self.config.max_action_dim - action.shape[-1]
                 action = torch.cat(
                     [action, torch.zeros(action.shape[0], action.shape[1], pad_size, device=action.device)],
                     dim=-1,
@@ -1902,6 +1965,10 @@ class WallXPolicy(PreTrainedPolicy):
                     ],
                     dim=-1,
                 )
+            elif action.shape[-1] > self.config.max_action_dim:
+                raise ValueError(
+                    f"Action dimension {action.shape[-1]} exceeds max_action_dim {self.config.max_action_dim}"
+                )
         else:
             action_dim = self.config.output_features[ACTION].shape[0]
             dof_mask = torch.cat(
@@ -1910,7 +1977,10 @@ class WallXPolicy(PreTrainedPolicy):
                         batch_size, self.config.chunk_size, action_dim, device=batch[OBS_STATE].device
                     ),
                     torch.zeros(
-                        batch_size, self.config.chunk_size, 20 - action_dim, device=batch[OBS_STATE].device
+                        batch_size,
+                        self.config.chunk_size,
+                        self.config.max_action_dim - action_dim,
+                        device=batch[OBS_STATE].device,
                     ),
                 ],
                 dim=-1,
@@ -1930,12 +2000,26 @@ class WallXPolicy(PreTrainedPolicy):
             text=all_texts,
             images=all_image_inputs,
             videos=None,
+            device=batch[OBS_STATE].device,
             padding=True,
             truncation=True,
             return_tensors="pt",
             max_length=TOKENIZER_MAX_LENGTH,
         )
 
+        if compute_position_ids:
+            # Qwen's RoPE indexing uses Python list/scalar conversions. Run it while the
+            # tokenizer and grid metadata are still on CPU, then move the compact result.
+            position_ids, rope_deltas = self.model.get_rope_index(
+                inputs.input_ids,
+                inputs.get("image_grid_thw"),
+                inputs.get("video_grid_thw"),
+                inputs.get("second_per_grid_ts"),
+                inputs.attention_mask,
+            )
+            inputs["position_ids"] = position_ids
+            inputs["rope_deltas"] = rope_deltas
+
         # ==================== ADDITIONAL INPUTS ====================
         action_token_id = self.model.processor.tokenizer.convert_tokens_to_ids("<|action|>")
         moe_token_types = inputs.input_ids == action_token_id
@@ -1952,7 +2036,7 @@ class WallXPolicy(PreTrainedPolicy):
         )
 
         # Move all tensors to the correct device
-        device = self.config.device
+        device = batch[OBS_STATE].device
         for key, value in inputs.items():
             if isinstance(value, torch.Tensor):
                 inputs[key] = value.to(device)
@@ -1972,9 +2056,7 @@ class WallXPolicy(PreTrainedPolicy):
         Returns:
             tuple: (loss, loss_dict)
         """
-        batch = self.preprocess_inputs(
-            batch,
-        )
+        batch = self.preprocess_inputs(batch, compute_position_ids=True)
 
         # Call the underlying model's forward with mode="train"
         outputs = self.model(**batch, mode="train")
@@ -1982,19 +2064,19 @@ class WallXPolicy(PreTrainedPolicy):
         # Extract losses from output
         loss = outputs.loss
         loss_dict = {
-            "loss": loss.item() if loss is not None else 0.0,
+            "loss": loss.detach() if loss is not None else 0.0,
         }
 
         if outputs.flow_loss is not None:
-            loss_dict["flow_loss"] = outputs.flow_loss.item()
+            loss_dict["flow_loss"] = outputs.flow_loss.detach()
         if outputs.cross_entropy_loss is not None:
-            loss_dict["cross_entropy_loss"] = outputs.cross_entropy_loss.item()
+            loss_dict["cross_entropy_loss"] = outputs.cross_entropy_loss.detach()
 
         # Add channel losses if available
         if outputs.channel_loss_dict is not None:
             for key, value in outputs.channel_loss_dict.items():
                 if isinstance(value, torch.Tensor):
-                    loss_dict[f"channel_{key}"] = value.item()
+                    loss_dict[f"channel_{key}"] = value.detach()
 
         return loss, loss_dict
 

src/lerobot/policies/wall_x/processor_wall_x.py

@@ -20,19 +20,13 @@ import torch
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
     ComplementaryDataProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStepRegistry,
-    RenameObservationsProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
-from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
 
 from .configuration_wall_x import WallXConfig
 
@@ -65,37 +59,22 @@ def make_wall_x_pre_post_processors(
         A tuple containing the configured pre-processor and post-processor pipelines
     """
 
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
+        steps.rename_observations,
+        steps.add_batch_dim,
         WallXTaskProcessor(),  # Process task description
-        NormalizerProcessorStep(
-            features={**config.input_features, **config.output_features},
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-        DeviceProcessorStep(device=config.device),
+        steps.normalize,
+        steps.to_device,
     ]
 
     output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
-        DeviceProcessorStep(device="cpu"),
+        steps.unnormalize,
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)
 
 
 @ProcessorStepRegistry.register(name="wall_x_task_processor")

src/lerobot/policies/wall_x/qwen_model/__init__.py

@@ -0,0 +1,45 @@
+#!/usr/bin/env python
+
+# Copyright 2025 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.
+
+from .configuration_qwen2_5_vl import (
+    Qwen2_5_VLConfig,
+    Qwen2_5_VLTextConfig,
+    Qwen2_5_VLVisionConfig,
+)
+from .qwen2_5_vl_moe import (
+    BlockSparseMLP,
+    Qwen2_5_VLACausalLMOutputWithPast,
+    Qwen2_5_VLDecoderLayer_with_MoE,
+    Qwen2_5_VLMoEModel,
+    SparseMoeBlock,
+)
+from .vision_attention import (
+    WallXVisionAttention,
+    configure_wall_x_vision_attention,
+)
+
+__all__ = [
+    "BlockSparseMLP",
+    "Qwen2_5_VLACausalLMOutputWithPast",
+    "Qwen2_5_VLConfig",
+    "Qwen2_5_VLDecoderLayer_with_MoE",
+    "Qwen2_5_VLMoEModel",
+    "Qwen2_5_VLTextConfig",
+    "Qwen2_5_VLVisionConfig",
+    "SparseMoeBlock",
+    "WallXVisionAttention",
+    "configure_wall_x_vision_attention",
+]

src/lerobot/policies/wall_x/qwen_model/configuration_qwen2_5_vl.py

@@ -1,250 +1,114 @@
-from transformers.configuration_utils import PretrainedConfig
-from transformers.modeling_rope_utils import rope_config_validation
+#!/usr/bin/env python
+
+# Copyright 2025 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.
+
+"""Wall-X configuration extensions for the native Transformers Qwen2.5-VL config."""
+
+from dataclasses import dataclass
+from typing import TYPE_CHECKING
+
+from huggingface_hub.dataclasses import strict
+
+from lerobot.utils.import_utils import _transformers_available
+
+if TYPE_CHECKING or _transformers_available:
+    from transformers.models.qwen2_5_vl.configuration_qwen2_5_vl import (
+        Qwen2_5_VLConfig as TransformersQwen2_5_VLConfig,
+        Qwen2_5_VLTextConfig as TransformersQwen2_5_VLTextConfig,
+        Qwen2_5_VLVisionConfig,
+    )
+else:
+
+    @dataclass
+    class _TransformersConfigFallback:
+        """Import-safe stand-in used only when Transformers is unavailable."""
+
+    TransformersQwen2_5_VLConfig = _TransformersConfigFallback
+    TransformersQwen2_5_VLTextConfig = _TransformersConfigFallback
+    Qwen2_5_VLVisionConfig = None
+
+# Wall-X checkpoints pre0.6.0 use the legacy, flat Qwen2.5-VL config layout.  The native
+# ``Qwen2_5_VLConfig`` accepts that layout and moves text-model fields into its
+# ``text_config`` sub-config, so only the Wall-X-specific MoE fields need to be
+# declared here.
+_LEGACY_TEXT_ATTRIBUTES = {
+    "attention_dropout",
+    "attention_moe",
+    "dim_inputs",
+    "dof_config",
+    "experts",
+    "hidden_act",
+    "hidden_size",
+    "initializer_range",
+    "intermediate_size",
+    "layer_types",
+    "max_position_embeddings",
+    "max_window_layers",
+    "mlp_moe",
+    "noise_scheduler",
+    "num_attention_heads",
+    "num_experts",
+    "num_hidden_layers",
+    "num_key_value_heads",
+    "pad_token_id",
+    "rms_norm_eps",
+    "sliding_window",
+    "use_cache",
+    "use_sliding_window",
+    "vocab_size",
+}
 
 
-class Qwen2_5_VLVisionConfig(PretrainedConfig):
-    model_type = "qwen2_5_vl"
-    base_config_key = "vision_config"
+@strict
+class Qwen2_5_VLTextConfig(TransformersQwen2_5_VLTextConfig):  # noqa: N801
+    """Native Qwen2.5-VL text config plus Wall-X's hard-routed MoE settings."""
 
-    def __init__(
-        self,
-        depth=32,
-        hidden_size=3584,
-        hidden_act="silu",
-        intermediate_size=3420,
-        num_heads=16,
-        in_channels=3,
-        patch_size=14,
-        spatial_merge_size=2,
-        temporal_patch_size=2,
-        tokens_per_second=4,
-        window_size=112,
-        out_hidden_size=3584,
-        fullatt_block_indexes=[7, 15, 23, 31],
-        initializer_range=0.02,
-        **kwargs,
-    ):
-        super().__init__(**kwargs)
+    num_experts: int = 4
+    experts: list[dict] | None = None
+    dof_config: dict | None = None
+    noise_scheduler: dict | None = None
+    dim_inputs: tuple[int, ...] | list[int] = (1536, 1536)
+    attention_moe: bool = False
+    mlp_moe: bool = False
 
-        self.depth = depth
-        self.hidden_size = hidden_size
-        self.hidden_act = hidden_act
-        self.intermediate_size = intermediate_size
-        self.num_heads = num_heads
-        self.in_channels = in_channels
-        self.patch_size = patch_size
-        self.spatial_merge_size = spatial_merge_size
-        self.temporal_patch_size = temporal_patch_size
-        self.tokens_per_second = tokens_per_second
-        self.window_size = window_size
-        self.fullatt_block_indexes = fullatt_block_indexes
-        self.out_hidden_size = out_hidden_size
-        self.initializer_range = initializer_range
+    def __post_init__(self, **kwargs):
+        self.dim_inputs = tuple(self.dim_inputs)
+        super().__post_init__(**kwargs)
 
 
-class Qwen2_5_VLConfig(PretrainedConfig):
-    r"""
-    This is the configuration class to store the configuration of a [`Qwen2_5_VLModel`]. It is used to instantiate a
-    Qwen2-VL model according to the specified arguments, defining the model architecture. Instantiating a configuration
-    with the defaults will yield a similar configuration to that of
-    Qwen2-VL-7B-Instruct [Qwen/Qwen2-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct).
+@strict
+class Qwen2_5_VLConfig(TransformersQwen2_5_VLConfig):  # noqa: N801
+    """Native composite Qwen2.5-VL config with a Wall-X text sub-config.
 
-    Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
-    documentation from [`PretrainedConfig`] for more information.
+    The native composite loader supports both current nested configs and the
+    flat layout used by existing ``wall-oss-flow`` checkpoints.
+    """
 
-
-    Args:
-        vocab_size (`int`, *optional*, defaults to 152064):
-            Vocabulary size of the Qwen2_5_VL model. Defines the number of different tokens that can be represented by the
-            `inputs_ids` passed when calling [`Qwen2_5_VLModel`]
-        hidden_size (`int`, *optional*, defaults to 8192):
-            Dimension of the hidden representations.
-        intermediate_size (`int`, *optional*, defaults to 29568):
-            Dimension of the MLP representations.
-        num_hidden_layers (`int`, *optional*, defaults to 80):
-            Number of hidden layers in the Transformer encoder.
-        num_attention_heads (`int`, *optional*, defaults to 64):
-            Number of attention heads for each attention layer in the Transformer encoder.
-        num_key_value_heads (`int`, *optional*, defaults to 8):
-            This is the number of key_value heads that should be used to implement Grouped Query Attention. If
-            `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
-            `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
-            converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
-            by meanpooling all the original heads within that group. For more details checkout [this
-            paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
-        hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
-            The non-linear activation function (function or string) in the decoder.
-        max_position_embeddings (`int`, *optional*, defaults to 32768):
-            The maximum sequence length that this model might ever be used with.
-        initializer_range (`float`, *optional*, defaults to 0.02):
-            The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
-        rms_norm_eps (`float`, *optional*, defaults to 1e-05):
-            The epsilon used by the rms normalization layers.
-        use_cache (`bool`, *optional*, defaults to `True`):
-            Whether or not the model should return the last key/values attentions (not used by all models). Only
-            relevant if `config.is_decoder=True`.
-        tie_word_embeddings (`bool`, *optional*, defaults to `False`):
-            Whether the model's input and output word embeddings should be tied.
-        rope_theta (`float`, *optional*, defaults to 1000000.0):
-            The base period of the RoPE embeddings.
-        use_sliding_window (`bool`, *optional*, defaults to `False`):
-            Whether to use sliding window attention.
-        sliding_window (`int`, *optional*, defaults to 4096):
-            Sliding window attention (SWA) window size. If not specified, will default to `4096`.
-        max_window_layers (`int`, *optional*, defaults to 80):
-            The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
-        attention_dropout (`float`, *optional*, defaults to 0.0):
-            The dropout ratio for the attention probabilities.
-        vision_config (`Dict`, *optional*):
-            The config for the visual encoder initialization.
-        rope_scaling (`Dict`, *optional*):
-            Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
-            and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
-            accordingly.
-            Expected contents:
-                `rope_type` (`str`):
-                    The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
-                    'llama3'], with 'default' being the original RoPE implementation.
-                `factor` (`float`, *optional*):
-                    Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
-                    most scaling types, a `factor` of x will enable the model to handle sequences of length x *
-                    original maximum pre-trained length.
-                `original_max_position_embeddings` (`int`, *optional*):
-                    Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
-                    pretraining.
-                `attention_factor` (`float`, *optional*):
-                    Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
-                    computation. If unspecified, it defaults to value recommended by the implementation, using the
-                    `factor` field to infer the suggested value.
-                `beta_fast` (`float`, *optional*):
-                    Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
-                    ramp function. If unspecified, it defaults to 32.
-                `beta_slow` (`float`, *optional*):
-                    Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
-                    ramp function. If unspecified, it defaults to 1.
-                `short_factor` (`List[float]`, *optional*):
-                    Only used with 'longrope'. The scaling factor to be applied to short contexts (<
-                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
-                    size divided by the number of attention heads divided by 2
-                `long_factor` (`List[float]`, *optional*):
-                    Only used with 'longrope'. The scaling factor to be applied to long contexts (<
-                    `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
-                    size divided by the number of attention heads divided by 2
-                `low_freq_factor` (`float`, *optional*):
-                    Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
-                `high_freq_factor` (`float`, *optional*):
-                    Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
-
-    ```python
-    >>> from transformers import Qwen2_5_VLForConditionalGeneration, Qwen2_5_VLConfig
-
-    >>> # Initializing a Qwen2_5_VL style configuration
-    >>> configuration = Qwen2_5_VLConfig()
-
-    >>> # Initializing a model from the Qwen2-VL-7B style configuration
-    >>> model = Qwen2_5_VLForConditionalGeneration(configuration)
-
-    >>> # Accessing the model configuration
-    >>> configuration = model.config
-    ```"""
-
-    model_type = "qwen2_5_vl"
-    sub_configs = {"vision_config": Qwen2_5_VLVisionConfig}
-    keys_to_ignore_at_inference = ["past_key_values"]
-    # Default tensor parallel plan for base model `Qwen2_5_VL`
-    base_model_tp_plan = {
-        "layers.*.self_attn.q_proj": "colwise",
-        "layers.*.self_attn.k_proj": "colwise",
-        "layers.*.self_attn.v_proj": "colwise",
-        "layers.*.self_attn.o_proj": "rowwise",
-        "layers.*.mlp.gate_proj": "colwise",
-        "layers.*.mlp.up_proj": "colwise",
-        "layers.*.mlp.down_proj": "rowwise",
-    }
-    base_model_pp_plan = {
-        "embed_tokens": (["input_ids"], ["inputs_embeds"]),
-        "layers": (["hidden_states", "attention_mask"], ["hidden_states"]),
-        "norm": (["hidden_states"], ["hidden_states"]),
+    sub_configs = {
+        "vision_config": Qwen2_5_VLVisionConfig,
+        "text_config": Qwen2_5_VLTextConfig,
     }
 
-    def __init__(
-        self,
-        vocab_size=152064,
-        hidden_size=8192,
-        intermediate_size=29568,
-        num_hidden_layers=80,
-        num_attention_heads=64,
-        num_key_value_heads=8,
-        hidden_act="silu",
-        max_position_embeddings=32768,
-        initializer_range=0.02,
-        rms_norm_eps=1e-05,
-        use_cache=True,
-        tie_word_embeddings=False,
-        rope_theta=1000000.0,
-        use_sliding_window=False,
-        sliding_window=4096,
-        max_window_layers=80,
-        attention_dropout=0.0,
-        vision_config=None,
-        rope_scaling=None,
-        num_experts=4,
-        experts=None,
-        dof_config=None,
-        noise_scheduler=None,
-        dim_inputs=(1536, 1536),
-        attention_moe=False,
-        mlp_moe=False,
-        **kwargs,
-    ):
-        if isinstance(vision_config, dict):
-            self.vision_config = self.sub_configs["vision_config"](**vision_config)
-        elif vision_config is None:
-            self.vision_config = self.sub_configs["vision_config"]()
+    def __getattr__(self, name):
+        """Keep legacy direct access to fields now owned by ``text_config``.
 
-        self.vocab_size = vocab_size
-        self.max_position_embeddings = max_position_embeddings
-        self.hidden_size = hidden_size
-        self.intermediate_size = intermediate_size
-        self.num_hidden_layers = num_hidden_layers
-        self.num_attention_heads = num_attention_heads
-        self.use_sliding_window = use_sliding_window
-        self.sliding_window = sliding_window
-        self.max_window_layers = max_window_layers
-        self.layer_types = ["dense"] * num_hidden_layers
-
-        # for backward compatibility
-        if num_key_value_heads is None:
-            num_key_value_heads = num_attention_heads
-
-        self.num_key_value_heads = num_key_value_heads
-        self.hidden_act = hidden_act
-        self.initializer_range = initializer_range
-        self.rms_norm_eps = rms_norm_eps
-        self.use_cache = use_cache
-        self.rope_theta = rope_theta
-        self.attention_dropout = attention_dropout
-        self.rope_scaling = rope_scaling
-
-        self.num_experts = num_experts
-        self.experts = experts
-        self.dof_config = dof_config
-        self.noise_scheduler = noise_scheduler
-        self.dim_inputs = tuple(dim_inputs)
-        self.attention_moe = attention_moe
-        self.mlp_moe = mlp_moe
-
-        if self.rope_scaling is not None and "type" in self.rope_scaling:
-            if self.rope_scaling["type"] == "mrope":
-                self.rope_scaling["type"] = "default"
-            self.rope_scaling["rope_type"] = self.rope_scaling["type"]
-        rope_config_validation(self, ignore_keys={"mrope_section"})
-
-        super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
-
-    @property
-    def text_config(self):
-        return self
-
-
-__all__ = ["Qwen2_5_VLConfig"]
+        Wall-X historically used a flat config and accesses fields such as
+        ``hidden_size`` and ``num_experts`` directly. Forwarding unknown
+        attributes preserves that API without duplicating the native config.
+        """
+        text_config = self.__dict__.get("text_config")
+        if name in _LEGACY_TEXT_ATTRIBUTES and text_config is not None and hasattr(text_config, name):
+            return getattr(text_config, name)
+        raise AttributeError(f"{type(self).__name__!s} has no attribute {name!r}")

src/lerobot/policies/wall_x/qwen_model/vision_attention.py

@@ -0,0 +1,208 @@
+#!/usr/bin/env python
+
+# Copyright 2025 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.
+
+"""Wall-X vision attention backends.
+
+Qwen2.5-VL's native non-Flash vision path splits a packed image sequence into
+Python-level chunks before calling attention. Wall-X batches many camera frames,
+so that path launches thousands of tiny attention operations per training step.
+This module keeps the native SDPA path as a portable fallback and adds a packed
+``torch.nn.attention.varlen`` path that consumes Qwen's existing ``cu_seqlens``
+metadata directly.
+"""
+
+from __future__ import annotations
+
+import inspect
+import logging
+from functools import lru_cache
+from typing import TYPE_CHECKING, Literal
+
+import torch
+import torch.nn as nn
+
+from lerobot.utils.import_utils import _transformers_available
+
+if TYPE_CHECKING or _transformers_available:
+    from transformers.models.qwen2_5_vl.modeling_qwen2_5_vl import (
+        Qwen2_5_VLVisionAttention,
+        apply_rotary_pos_emb_vision,
+    )
+else:
+    Qwen2_5_VLVisionAttention = nn.Module
+    apply_rotary_pos_emb_vision = None
+
+try:
+    from torch.nn.attention.varlen import varlen_attn as _varlen_attn
+except ImportError:  # torch<2.10
+    _varlen_attn = None
+
+_VARLEN_USES_WINDOW_SIZE = (
+    _varlen_attn is not None and "window_size" in inspect.signature(_varlen_attn).parameters
+)
+
+
+VisionAttentionBackend = Literal["auto", "sdpa", "varlen"]
+
+logger = logging.getLogger(__name__)
+
+
+@lru_cache
+def _log_resolved_backend(requested: str, resolved: str) -> None:
+    logger.info("Wall-X vision attention backend: %s (requested: %s)", resolved, requested)
+
+
+def _varlen_unavailable_reason(
+    hidden_states: torch.Tensor,
+    position_embeddings: tuple[torch.Tensor, torch.Tensor] | None,
+) -> str | None:
+    if _varlen_attn is None:
+        return "torch.nn.attention.varlen is unavailable (PyTorch 2.10 or newer is required)"
+    if position_embeddings is None:
+        return "precomputed vision position embeddings were not provided"
+    if hidden_states.device.type != "cuda" or torch.version.cuda is None:
+        return "packed varlen attention requires an NVIDIA CUDA device"
+    if hidden_states.dtype not in {torch.float16, torch.bfloat16}:
+        return f"packed varlen attention requires float16 or bfloat16 inputs, got {hidden_states.dtype}"
+    major, _minor = torch.cuda.get_device_capability(hidden_states.device)
+    if major < 8:
+        return "packed varlen attention requires an NVIDIA Ampere GPU or newer"
+    return None
+
+
+def _supports_varlen_attention(
+    hidden_states: torch.Tensor,
+    position_embeddings: tuple[torch.Tensor, torch.Tensor] | None,
+) -> bool:
+    return _varlen_unavailable_reason(hidden_states, position_embeddings) is None
+
+
+class WallXVisionAttention(Qwen2_5_VLVisionAttention):
+    """Qwen2.5-VL vision attention with packed varlen and native SDPA fallback."""
+
+    def __init__(self, config, backend: VisionAttentionBackend):
+        super().__init__(config)
+        self.wallx_backend = backend
+        self._resolved_backend_key = None
+        self._resolved_backend = None
+
+    def _resolve_backend(
+        self,
+        hidden_states: torch.Tensor,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None,
+    ) -> str:
+        key = (
+            hidden_states.device.type,
+            hidden_states.device.index,
+            hidden_states.dtype,
+            position_embeddings is not None,
+        )
+        if self._resolved_backend_key == key:
+            return self._resolved_backend
+
+        use_varlen = self.wallx_backend != "sdpa" and _supports_varlen_attention(
+            hidden_states, position_embeddings
+        )
+        if self.wallx_backend == "varlen" and not use_varlen:
+            reason = _varlen_unavailable_reason(hidden_states, position_embeddings)
+            raise RuntimeError(f"Wall-X vision_attn_implementation='varlen' cannot be used: {reason}")
+
+        resolved_backend = "varlen" if use_varlen else "sdpa"
+        self._resolved_backend_key = key
+        self._resolved_backend = resolved_backend
+        _log_resolved_backend(self.wallx_backend, resolved_backend)
+        return resolved_backend
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        rotary_pos_emb: torch.Tensor | None = None,
+        position_embeddings: tuple[torch.Tensor, torch.Tensor] | None = None,
+        **kwargs,
+    ) -> torch.Tensor:
+        del rotary_pos_emb
+
+        if self._resolve_backend(hidden_states, position_embeddings) == "sdpa":
+            return super().forward(
+                hidden_states=hidden_states,
+                cu_seqlens=cu_seqlens,
+                position_embeddings=position_embeddings,
+                **kwargs,
+            )
+
+        seq_length = hidden_states.shape[0]
+        query_states, key_states, value_states = (
+            self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
+        )
+
+        cos, sin = position_embeddings
+        query_states, key_states = apply_rotary_pos_emb_vision(
+            query_states,
+            key_states,
+            cos,
+            sin,
+        )
+
+        if cu_seqlens.dtype != torch.int32:
+            cu_seqlens = cu_seqlens.to(dtype=torch.int32)
+        max_seqlen = int((cu_seqlens[1:] - cu_seqlens[:-1]).max().item())
+        varlen_kwargs = {"scale": self.scaling}
+        if _VARLEN_USES_WINDOW_SIZE:
+            varlen_kwargs["window_size"] = (-1, -1)
+        else:  # Stable PyTorch 2.10 API; pre-release variants used window_size.
+            varlen_kwargs["is_causal"] = False
+        attn_output = _varlen_attn(
+            query_states,
+            key_states,
+            value_states,
+            cu_seqlens,
+            cu_seqlens,
+            max_seqlen,
+            max_seqlen,
+            **varlen_kwargs,
+        )
+        attn_output = attn_output.reshape(seq_length, -1).contiguous()
+        return self.proj(attn_output)
+
+
+def configure_wall_x_vision_attention(
+    vision_model: nn.Module,
+    backend: VisionAttentionBackend,
+) -> None:
+    """Install Wall-X's scoped packed attention without changing checkpoint keys."""
+    if backend == "sdpa":
+        _log_resolved_backend(backend, "sdpa")
+        return
+    if backend == "varlen" and _varlen_attn is None:
+        raise RuntimeError(
+            "Wall-X vision_attn_implementation='varlen' requires torch.nn.attention.varlen "
+            "from PyTorch 2.10 or newer"
+        )
+    if backend == "auto" and _varlen_attn is None:
+        _log_resolved_backend(backend, "sdpa")
+        return
+
+    for block in vision_model.blocks:
+        previous_attention = block.attn
+        replacement = WallXVisionAttention(previous_attention.config, backend=backend)
+        replacement.to(
+            device=previous_attention.qkv.weight.device,
+            dtype=previous_attention.qkv.weight.dtype,
+        )
+        replacement.load_state_dict(previous_attention.state_dict(), strict=True)
+        replacement.train(previous_attention.training)
+        block.attn = replacement

src/lerobot/policies/wall_x/utils.py

@@ -116,6 +116,7 @@ def preprocesser_call(
     images: list | Any | None = None,
     text: str | list[str] | None = None,
     videos: list | Any | None = None,
+    device: torch.device | str | None = None,
     padding: bool | str = False,
     truncation: bool | None = None,
     max_length: int | None = None,
@@ -134,6 +135,7 @@ def preprocesser_call(
         images: Input images (PIL, numpy arrays, or torch tensors)
         text: Text or list of texts to tokenize
         videos: Input videos (numpy arrays or torch tensors)
+        device: Device on which image/video preprocessing should run
         padding: Whether to pad sequences to same length
         truncation: Whether to truncate sequences longer than max_length
         max_length: Maximum length for truncation/padding
@@ -151,7 +153,11 @@ def preprocesser_call(
     """
     # Process image inputs
     if images is not None and len(images) > 0:
-        image_inputs = processor.image_processor(images=images, return_tensors=return_tensors)
+        image_inputs = processor.image_processor(
+            images=images,
+            return_tensors=return_tensors,
+            device=device,
+        )
         image_grid_thw = image_inputs["image_grid_thw"]
     else:
         image_inputs = {}
@@ -159,7 +165,11 @@ def preprocesser_call(
 
     # Process video inputs
     if videos is not None:
-        videos_inputs = processor.image_processor(videos=videos, return_tensors=return_tensors)
+        videos_inputs = processor.image_processor(
+            videos=videos,
+            return_tensors=return_tensors,
+            device=device,
+        )
         video_grid_thw = videos_inputs["video_grid_thw"]
     else:
         videos_inputs = {}
@@ -413,10 +423,7 @@ def get_task_instruction(
         }
     )
 
-    if priority_order is not None:
-        priority_order = OrderedDict(priority_order)
-    else:
-        priority_order = default_priority_order
+    priority_order = OrderedDict(priority_order) if priority_order is not None else default_priority_order
 
     got_instruction = False
     task_instruction = ""
@@ -424,9 +431,8 @@ def get_task_instruction(
     # Sample instruction components based on priority probabilities
     for key, prob in priority_order.items():
         if key in frame_instruction_info and frame_instruction_info[key] != "":
-            if got_instruction:
-                if random.random() >= prob:
-                    continue
+            if got_instruction and random.random() >= prob:
+                continue
 
             task_instruction += f"\n{frame_instruction_info[key]}"
             got_instruction = True
@@ -538,10 +544,7 @@ def img_key_mapping(img_keys: list[str]) -> list[str]:
         if key in CAMERA_NAME_MAPPING:
             key = CAMERA_NAME_MAPPING[key]
         else:
-            if "view" in key:
-                key = key.replace("_", " ")
-            else:
-                key = key + " view"
+            key = key.replace("_", " ") if "view" in key else key + " view"
         processed_img_keys.append(key)
     return processed_img_keys
 

src/lerobot/policies/xvla/processor_xvla.py

@@ -22,19 +22,14 @@ import torch
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
 from lerobot.processor import (
-    AddBatchDimensionProcessorStep,
-    DeviceProcessorStep,
-    NormalizerProcessorStep,
     ObservationProcessorStep,
     PolicyAction,
     PolicyProcessorPipeline,
     ProcessorStep,
     ProcessorStepRegistry,
-    RenameObservationsProcessorStep,
     TokenizerProcessorStep,
-    UnnormalizerProcessorStep,
-    policy_action_to_transition,
-    transition_to_policy_action,
+    make_default_policy_processor_steps,
+    make_policy_processor_pipelines,
 )
 from lerobot.types import EnvTransition, TransitionKey
 from lerobot.utils.constants import (
@@ -42,8 +37,6 @@ from lerobot.utils.constants import (
     OBS_IMAGES,
     OBS_PREFIX,
     OBS_STATE,
-    POLICY_POSTPROCESSOR_DEFAULT_NAME,
-    POLICY_PREPROCESSOR_DEFAULT_NAME,
 )
 
 from .configuration_xvla import XVLAConfig
@@ -61,10 +54,11 @@ def make_xvla_pre_post_processors(
     Build the LeRobot processor pipelines for XVLA.
     """
 
-    features = {**config.input_features, **config.output_features}
+    steps = make_default_policy_processor_steps(config, dataset_stats)
+
     input_steps = [
-        RenameObservationsProcessorStep(rename_map={}),
-        AddBatchDimensionProcessorStep(),
+        steps.rename_observations,
+        steps.add_batch_dim,
         TokenizerProcessorStep(
             tokenizer_name=config.tokenizer_name,
             max_length=config.tokenizer_max_length,
@@ -74,32 +68,15 @@ def make_xvla_pre_post_processors(
         XVLAImageToFloatProcessorStep(),
         XVLAImageNetNormalizeProcessorStep(),
         XVLAAddDomainIdProcessorStep(),
-        DeviceProcessorStep(device=config.device),
-        NormalizerProcessorStep(
-            features=features, norm_map=config.normalization_mapping, stats=dataset_stats
-        ),
+        steps.to_device,
+        steps.normalize,
     ]
     output_steps = [
-        UnnormalizerProcessorStep(
-            features=config.output_features,
-            norm_map=config.normalization_mapping,
-            stats=dataset_stats,
-        ),
-        DeviceProcessorStep(device="cpu"),
+        steps.unnormalize,
+        steps.to_cpu,
     ]
 
-    return (
-        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
-            steps=input_steps,
-            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
-        ),
-        PolicyProcessorPipeline[PolicyAction, PolicyAction](
-            steps=output_steps,
-            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
-            to_transition=policy_action_to_transition,
-            to_output=transition_to_policy_action,
-        ),
-    )
+    return make_policy_processor_pipelines(input_steps=input_steps, output_steps=output_steps)
 
 
 # Custom XVLA processor steps

src/lerobot/processor/__init__.py

@@ -42,10 +42,14 @@ from .delta_action_processor import MapDeltaActionToRobotActionStep, MapTensorTo
 from .device_processor import DeviceProcessorStep
 from .env_processor import IsaaclabArenaProcessorStep, LiberoProcessorStep
 from .factory import (
+    DefaultPolicyProcessorSteps,
+    make_default_policy_processor_steps,
+    make_default_pre_post_processors,
     make_default_processors,
     make_default_robot_action_processor,
     make_default_robot_observation_processor,
     make_default_teleop_action_processor,
+    make_policy_processor_pipelines,
 )
 from .gym_action_processor import (
     Numpy2TorchActionProcessorStep,
@@ -129,10 +133,14 @@ __all__ = [
     "ImageCropResizeProcessorStep",
     "InfoProcessorStep",
     "InterventionActionProcessorStep",
+    "DefaultPolicyProcessorSteps",
+    "make_default_policy_processor_steps",
+    "make_default_pre_post_processors",
     "make_default_processors",
     "make_default_teleop_action_processor",
     "make_default_robot_action_processor",
     "make_default_robot_observation_processor",
+    "make_policy_processor_pipelines",
     "AbsoluteActionsProcessorStep",
     "RelativeActionsProcessorStep",
     "MapDeltaActionToRobotActionStep",

src/lerobot/processor/factory.py

@@ -14,15 +14,33 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from lerobot.types import RobotAction, RobotObservation
+from dataclasses import dataclass
+from typing import Any
 
+import torch
+
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.types import PolicyAction, RobotAction, RobotObservation
+from lerobot.utils.constants import POLICY_POSTPROCESSOR_DEFAULT_NAME, POLICY_PREPROCESSOR_DEFAULT_NAME
+
+from .batch_processor import AddBatchDimensionProcessorStep
 from .converters import (
     observation_to_transition,
+    policy_action_to_transition,
     robot_action_observation_to_transition,
     transition_to_observation,
+    transition_to_policy_action,
     transition_to_robot_action,
 )
-from .pipeline import IdentityProcessorStep, RobotProcessorPipeline
+from .device_processor import DeviceProcessorStep
+from .normalize_processor import NormalizerProcessorStep, UnnormalizerProcessorStep
+from .pipeline import (
+    IdentityProcessorStep,
+    PolicyProcessorPipeline,
+    ProcessorStep,
+    RobotProcessorPipeline,
+)
+from .rename_processor import RenameObservationsProcessorStep
 
 
 def make_default_teleop_action_processor() -> RobotProcessorPipeline[
@@ -61,3 +79,97 @@ def make_default_processors():
     robot_action_processor = make_default_robot_action_processor()
     robot_observation_processor = make_default_robot_observation_processor()
     return (teleop_action_processor, robot_action_processor, robot_observation_processor)
+
+
+@dataclass
+class DefaultPolicyProcessorSteps:
+    """The canonical processor steps shared by most policies' pre/post pipelines.
+
+    Policies compose these in their own order (step ORDER is a Hub-serialized contract
+    and intentionally stays explicit per policy) and interleave their custom steps.
+    """
+
+    rename_observations: RenameObservationsProcessorStep
+    add_batch_dim: AddBatchDimensionProcessorStep
+    to_device: DeviceProcessorStep
+    normalize: NormalizerProcessorStep
+    unnormalize: UnnormalizerProcessorStep
+    to_cpu: DeviceProcessorStep
+
+
+def make_default_policy_processor_steps(
+    config: PreTrainedConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    *,
+    normalizer_device: torch.device | str | None = None,
+) -> DefaultPolicyProcessorSteps:
+    """Construct the canonical policy processor steps from a policy config.
+
+    Args:
+        config: A `PreTrainedConfig` providing `device`, `input_features`,
+            `output_features` and `normalization_mapping`.
+        dataset_stats: Dataset statistics used for (un)normalization.
+        normalizer_device: Device passed to `NormalizerProcessorStep` (some policies pin
+            their normalization stats to the policy device; most leave it unset).
+    """
+    return DefaultPolicyProcessorSteps(
+        rename_observations=RenameObservationsProcessorStep(rename_map={}),
+        add_batch_dim=AddBatchDimensionProcessorStep(),
+        to_device=DeviceProcessorStep(device=config.device),
+        normalize=NormalizerProcessorStep(
+            features={**config.input_features, **config.output_features},
+            norm_map=config.normalization_mapping,
+            stats=dataset_stats,
+            device=normalizer_device,
+        ),
+        unnormalize=UnnormalizerProcessorStep(
+            features=config.output_features, norm_map=config.normalization_mapping, stats=dataset_stats
+        ),
+        to_cpu=DeviceProcessorStep(device="cpu"),
+    )
+
+
+def make_policy_processor_pipelines(
+    input_steps: list[ProcessorStep],
+    output_steps: list[ProcessorStep],
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """Wrap pre/post step lists into the canonical policy pipeline pair.
+
+    Uses the standard pipeline names (which determine the serialized JSON filenames on
+    the Hub) and the standard policy-action converters on the postprocessor.
+    """
+    return (
+        PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](
+            steps=input_steps,
+            name=POLICY_PREPROCESSOR_DEFAULT_NAME,
+        ),
+        PolicyProcessorPipeline[PolicyAction, PolicyAction](
+            steps=output_steps,
+            name=POLICY_POSTPROCESSOR_DEFAULT_NAME,
+            to_transition=policy_action_to_transition,
+            to_output=transition_to_policy_action,
+        ),
+    )
+
+
+def make_default_pre_post_processors(
+    config: PreTrainedConfig,
+    dataset_stats: dict[str, dict[str, torch.Tensor]] | None = None,
+    *,
+    normalizer_device: torch.device | str | None = None,
+) -> tuple[
+    PolicyProcessorPipeline[dict[str, Any], dict[str, Any]],
+    PolicyProcessorPipeline[PolicyAction, PolicyAction],
+]:
+    """The pure-scaffold policy pipeline pair: Rename -> Batch -> Device -> Normalize,
+    and Unnormalize -> Device(cpu). Policies with custom steps or a different step order
+    compose `make_default_policy_processor_steps` themselves instead.
+    """
+    s = make_default_policy_processor_steps(config, dataset_stats, normalizer_device=normalizer_device)
+    return make_policy_processor_pipelines(
+        input_steps=[s.rename_observations, s.add_batch_dim, s.to_device, s.normalize],
+        output_steps=[s.unnormalize, s.to_cpu],
+    )

tests/policies/common/test_flow_matching.py

@@ -0,0 +1,193 @@
+#!/usr/bin/env python
+
+# Copyright 2026 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.
+
+"""Behavior-pinning tests for the shared flow-matching sampling primitives.
+
+``euler_integrate`` is compared against a verbatim copy of the historical pi0/pi05/
+smolvla sampling loop (including its RTC hook semantics): any divergence from that
+reference is a behavior change for released checkpoints.
+"""
+
+import torch
+
+from lerobot.policies.common.flow_matching import (
+    euler_integrate,
+    sample_beta,
+    sample_noise,
+    sample_time_beta,
+)
+
+
+def test_sample_beta_range_dtype_and_reproducibility():
+    torch.manual_seed(0)
+    s1 = sample_beta(1.5, 1.0, 4096, "cpu")
+    torch.manual_seed(0)
+    s2 = sample_beta(1.5, 1.0, 4096, "cpu")
+    assert torch.equal(s1, s2)
+    assert s1.shape == (4096,) and s1.dtype == torch.float32
+    assert s1.min() >= 0.0 and s1.max() <= 1.0
+    # Beta(1.5, 1.0) mean is 1.5/2.5 = 0.6.
+    assert abs(s1.mean().item() - 0.6) < 0.02
+
+
+def test_sample_time_beta_openpi_convention():
+    torch.manual_seed(1)
+    time = sample_time_beta(4096, "cpu", alpha=1.5, beta=1.0, scale=0.999, offset=0.001)
+    assert time.dtype == torch.float32
+    assert time.min() >= 0.001 and time.max() <= 1.0
+    # Exact composition: Beta sample * scale + offset, same RNG stream.
+    torch.manual_seed(1)
+    expected = sample_beta(1.5, 1.0, 4096, "cpu") * 0.999 + 0.001
+    torch.testing.assert_close(time, expected, rtol=0, atol=0)
+
+
+def test_sample_noise_seeded():
+    torch.manual_seed(2)
+    n1 = sample_noise((2, 8, 4), "cpu")
+    torch.manual_seed(2)
+    n2 = sample_noise((2, 8, 4), "cpu")
+    assert torch.equal(n1, n2)
+    assert n1.dtype == torch.float32 and n1.shape == (2, 8, 4)
+
+
+def test_euler_integrate_constant_velocity_is_exact():
+    # With v_t == c constant, x_0 = x_1 + sum(dt * c) = x_1 - c exactly (num_steps * dt = -1).
+    noise = torch.randn(3, 5, 2)
+    c = torch.randn(3, 5, 2)
+    out = euler_integrate(lambda x_t, time: c, noise, num_steps=10)
+    torch.testing.assert_close(out, noise - c, rtol=0, atol=1e-6)
+
+
+def _reference_pi0_loop(denoise_fn, noise, num_steps, rtc_enabled, rtc_processor, kw):
+    """Verbatim structure of the historical pi0/pi05/smolvla sample_actions loop."""
+    bsize = noise.shape[0]
+    device = noise.device
+    dt = -1.0 / num_steps
+    x_t = noise
+    for step in range(num_steps):
+        time = 1.0 + step * dt
+        time_tensor = torch.tensor(time, dtype=torch.float32, device=device).expand(bsize)
+
+        def denoise_step_partial_call(input_x_t, current_timestep=time_tensor):
+            return denoise_fn(input_x_t, current_timestep)
+
+        if rtc_enabled:
+            v_t = rtc_processor.denoise_step(
+                x_t=x_t,
+                prev_chunk_left_over=kw.get("prev_chunk_left_over"),
+                inference_delay=kw.get("inference_delay"),
+                time=time,
+                original_denoise_step_partial=denoise_step_partial_call,
+                execution_horizon=kw.get("execution_horizon"),
+            )
+        else:
+            v_t = denoise_step_partial_call(x_t)
+        x_t = x_t + dt * v_t
+        if rtc_processor is not None and rtc_processor.is_debug_enabled():
+            rtc_processor.track(time=time, x_t=x_t, v_t=v_t)
+    return x_t
+
+
+class _StubRTCProcessor:
+    def __init__(self, debug_enabled: bool):
+        self._debug = debug_enabled
+        self.tracked = []
+        self.guidance_calls = []
+
+    def is_debug_enabled(self):
+        return self._debug
+
+    def denoise_step(
+        self,
+        x_t,
+        prev_chunk_left_over,
+        inference_delay,
+        time,
+        original_denoise_step_partial,
+        execution_horizon,
+    ):
+        self.guidance_calls.append(
+            {
+                "time": time,
+                "inference_delay": inference_delay,
+                "execution_horizon": execution_horizon,
+                "x_t": x_t.clone(),
+            }
+        )
+        return original_denoise_step_partial(x_t) * 0.5
+
+    def track(self, time, x_t, v_t):
+        self.tracked.append({"time": time, "x_t": x_t.clone(), "v_t": v_t.clone()})
+
+
+def _make_denoise_fn():
+    weight = torch.randn(4, 4) * 0.1
+
+    def denoise_fn(x_t, time_tensor):
+        return x_t @ weight + time_tensor[:, None, None]
+
+    return denoise_fn
+
+
+def test_euler_integrate_matches_historical_loop():
+    torch.manual_seed(3)
+    denoise_fn = _make_denoise_fn()
+    noise = torch.randn(2, 6, 4)
+    ref = _reference_pi0_loop(denoise_fn, noise, 10, rtc_enabled=False, rtc_processor=None, kw={})
+    out = euler_integrate(denoise_fn, noise, 10)
+    assert torch.equal(out, ref)
+
+
+def test_euler_integrate_rtc_guidance_and_kwarg_forwarding():
+    torch.manual_seed(4)
+    denoise_fn = _make_denoise_fn()
+    noise = torch.randn(2, 6, 4)
+    leftover = torch.randn(2, 6, 4)
+    kw = {"inference_delay": 3, "prev_chunk_left_over": leftover, "execution_horizon": 25}
+
+    ref_proc, new_proc = _StubRTCProcessor(False), _StubRTCProcessor(False)
+    ref = _reference_pi0_loop(denoise_fn, noise, 6, rtc_enabled=True, rtc_processor=ref_proc, kw=kw)
+    out = euler_integrate(
+        denoise_fn,
+        noise,
+        6,
+        rtc_processor=new_proc,
+        rtc_enabled=True,
+        inference_delay=3,
+        prev_chunk_left_over=leftover,
+        execution_horizon=25,
+    )
+    assert torch.equal(out, ref)
+    assert len(new_proc.guidance_calls) == 6
+    for ref_call, new_call in zip(ref_proc.guidance_calls, new_proc.guidance_calls, strict=True):
+        assert ref_call["time"] == new_call["time"]
+        assert new_call["inference_delay"] == 3 and new_call["execution_horizon"] == 25
+        # Guidance sees the PRE-update x_t.
+        assert torch.equal(ref_call["x_t"], new_call["x_t"])
+
+
+def test_euler_integrate_debug_tracking_fires_even_when_rtc_disabled():
+    # Historical behavior: track() fires whenever the processor exists and has debugging
+    # enabled, independent of whether RTC guidance is active.
+    torch.manual_seed(5)
+    denoise_fn = _make_denoise_fn()
+    noise = torch.randn(2, 6, 4)
+    proc = _StubRTCProcessor(True)
+    out = euler_integrate(denoise_fn, noise, 4, rtc_processor=proc, rtc_enabled=False)
+    assert len(proc.guidance_calls) == 0
+    assert len(proc.tracked) == 4
+    # track() receives the POST-update x_t; the last one is the returned sample.
+    assert torch.equal(proc.tracked[-1]["x_t"], out)

tests/policies/common/test_vla_utils.py

@@ -0,0 +1,289 @@
+#!/usr/bin/env python
+
+# Copyright 2026 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.
+
+"""Behavior-pinning tests for the shared VLA helpers.
+
+These helpers are the canonical versions of functions that used to be copy-pasted across
+the openpi-derived policies (pi0, pi05, pi0_fast, smolvla, eo1, xvla). The expected
+values below encode the historical per-policy behavior exactly; a failure here means a
+behavior change that would silently affect released checkpoints.
+"""
+
+import math
+
+import pytest
+import torch
+import torch.nn.functional as F  # noqa: N812
+
+from lerobot.policies.common.vla_utils import (
+    create_sinusoidal_pos_embedding,
+    fuse_action_time_embedding,
+    make_att_2d_masks,
+    pad_vector,
+    prepare_attention_masks_4d,
+    resize_with_pad,
+    resize_with_pad_torch,
+)
+from lerobot.utils.constants import OPENPI_ATTENTION_MASK_VALUE
+
+
+def test_create_sinusoidal_pos_embedding_matches_openpi_formula():
+    time = torch.tensor([0.0, 0.25, 1.0])
+    dim, min_period, max_period = 8, 4e-3, 4.0
+    emb = create_sinusoidal_pos_embedding(time, dim, min_period, max_period, device=torch.device("cpu"))
+
+    assert emb.shape == (3, dim)
+    # Independent recomputation of the openpi formula in float64.
+    fraction = torch.linspace(0.0, 1.0, dim // 2, dtype=torch.float64)
+    period = min_period * (max_period / min_period) ** fraction
+    scaling = 1.0 / period * 2 * math.pi
+    sin_input = scaling[None, :] * time.to(torch.float64)[:, None]
+    expected = torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
+    torch.testing.assert_close(emb, expected, rtol=1e-9, atol=1e-9)
+
+
+def test_create_sinusoidal_pos_embedding_validation():
+    with pytest.raises(ValueError, match="divisible by 2"):
+        create_sinusoidal_pos_embedding(torch.zeros(2), 7, 4e-3, 4.0, device=torch.device("cpu"))
+    with pytest.raises(ValueError, match="batch_size"):
+        create_sinusoidal_pos_embedding(torch.zeros(2, 2), 8, 4e-3, 4.0, device=torch.device("cpu"))
+
+
+def test_make_att_2d_masks_docstring_cases():
+    # Pure causal attention: [[1 1 1]]
+    pad = torch.ones(1, 3, dtype=torch.bool)
+    att = torch.tensor([[1, 1, 1]], dtype=torch.int32)
+    expected = torch.tensor([[[1, 0, 0], [1, 1, 0], [1, 1, 1]]], dtype=torch.bool)
+    assert torch.equal(make_att_2d_masks(pad, att), expected)
+
+    # Prefix-LM: [[0 0 1 1]] -> first two tokens attend bidirectionally, rest causal.
+    att = torch.tensor([[0, 0, 1, 1]], dtype=torch.int32)
+    pad = torch.ones(1, 4, dtype=torch.bool)
+    expected = torch.tensor([[[1, 1, 0, 0], [1, 1, 0, 0], [1, 1, 1, 0], [1, 1, 1, 1]]], dtype=torch.bool)
+    assert torch.equal(make_att_2d_masks(pad, att), expected)
+
+    # Padding removes rows and columns.
+    pad = torch.tensor([[True, True, False]])
+    att = torch.tensor([[0, 1, 1]], dtype=torch.int32)
+    out = make_att_2d_masks(pad, att)
+    assert not out[0, :, 2].any() and not out[0, 2, :].any()
+
+
+def test_make_att_2d_masks_validation():
+    with pytest.raises(ValueError):
+        make_att_2d_masks(torch.ones(3, dtype=torch.bool), torch.ones(1, 3, dtype=torch.int32))
+    with pytest.raises(ValueError):
+        make_att_2d_masks(torch.ones(1, 3, dtype=torch.bool), torch.ones(3, dtype=torch.int32))
+
+
+def test_prepare_attention_masks_4d():
+    masks = torch.tensor([[[True, False], [False, True]]])
+    out = prepare_attention_masks_4d(masks)
+    assert out.shape == (1, 1, 2, 2)
+    expected = torch.tensor([[[[0.0, OPENPI_ATTENTION_MASK_VALUE], [OPENPI_ATTENTION_MASK_VALUE, 0.0]]]])
+    assert torch.equal(out, expected)
+
+    out_bf16 = prepare_attention_masks_4d(masks, dtype=torch.bfloat16)
+    assert out_bf16.dtype == torch.bfloat16
+    assert torch.equal(out_bf16, expected.to(torch.bfloat16))
+
+
+def test_pad_vector_openpi_semantics():
+    v = torch.arange(6.0).reshape(2, 3)
+    padded = pad_vector(v, 5)
+    assert padded.shape == (2, 5)
+    assert torch.equal(padded[:, :3], v) and not padded[:, 3:].any()
+    # Already large enough (>=): returned unchanged, same object.
+    assert pad_vector(v, 3) is v
+    assert pad_vector(v, 2) is v
+    # 3D input.
+    v3 = torch.ones(2, 4, 3)
+    assert pad_vector(v3, 7).shape == (2, 4, 7)
+
+
+def test_pad_vector_truncate_semantics():
+    v = torch.arange(6.0).reshape(2, 3)
+    out = pad_vector(v, 2, truncate=True)
+    assert out.shape == (2, 2) and torch.equal(out, v[:, :2])
+    out = pad_vector(v, 5, truncate=True)
+    assert out.shape == (2, 5) and torch.equal(out[:, :3], v) and not out[:, 3:].any()
+    assert pad_vector(v, 0, truncate=True).shape == (2, 0)
+    assert pad_vector(v, 3, truncate=True) is v
+
+
+@pytest.mark.parametrize("channels_last", [True, False])
+def test_resize_with_pad_torch_centered(channels_last):
+    img = torch.rand(2, 3, 30, 60) if not channels_last else torch.rand(2, 30, 60, 3)
+    out = resize_with_pad_torch(img, 64, 64)
+    if channels_last:
+        assert out.shape == (2, 64, 64, 3)
+        # Aspect ratio preserved: 30x60 -> 32x64, padded 16 top and 16 bottom (centered).
+        assert not out[:, :16].any() and not out[:, -16:].any()
+        assert out[:, 16:48].abs().sum() > 0
+    else:
+        assert out.shape == (2, 3, 64, 64)
+        assert not out[:, :, :16].any() and not out[:, :, -16:].any()
+
+
+def test_resize_with_pad_torch_uint8_roundtrip():
+    img = (torch.rand(1, 3, 20, 20) * 255).to(torch.uint8)
+    out = resize_with_pad_torch(img, 40, 40)
+    assert out.dtype == torch.uint8 and out.shape == (1, 3, 40, 40)
+    with pytest.raises(ValueError, match="Unsupported image dtype"):
+        resize_with_pad_torch(torch.rand(1, 3, 8, 8, dtype=torch.float64), 16, 16)
+
+
+def test_resize_with_pad_top_left():
+    img = torch.rand(2, 3, 30, 60)
+    out = resize_with_pad(img, 64, 64, pad_value=-1.0)
+    assert out.shape == (2, 3, 64, 64)
+    # 30x60 -> 32x64; this variant pads on the TOP only (32 rows of pad_value).
+    assert torch.equal(out[:, :, :32], torch.full((2, 3, 32, 64), -1.0))
+    assert out[:, :, 32:].min() >= 0
+    # No-op fast path returns the same object.
+    assert resize_with_pad(img, 30, 60, pad_value=0.0) is img
+    with pytest.raises(ValueError, match="expected"):
+        resize_with_pad(torch.rand(3, 8, 8), 16, 16, pad_value=0.0)
+
+
+def test_clone_past_key_values():
+    pytest.importorskip("transformers")
+    from transformers import DynamicCache
+
+    from lerobot.policies.common.vla_utils import clone_past_key_values
+
+    cache = DynamicCache()
+    keys, values = torch.rand(1, 2, 4, 8), torch.rand(1, 2, 4, 8)
+    cache.update(keys, values, 0)
+    cloned = clone_past_key_values(cache)
+    (ck, cv, _), (ok, ov, _) = next(iter(cloned)), next(iter(cache))
+    assert torch.equal(ck, ok) and torch.equal(cv, ov)
+    # Deep copy: mutating the clone must not touch the original.
+    ck.zero_()
+    assert not torch.equal(ck, ok)
+
+
+def test_clone_past_key_values_is_fullgraph_compilable():
+    pytest.importorskip("transformers")
+    from transformers import DynamicCache
+
+    from lerobot.policies.common.vla_utils import clone_past_key_values
+
+    cache = DynamicCache()
+    keys, values = torch.rand(1, 2, 4, 8), torch.rand(1, 2, 4, 8)
+    cache.update(keys, values, 0)
+
+    compiled_clone = torch.compile(clone_past_key_values, backend="eager", fullgraph=True)
+    cloned = compiled_clone(cache)
+
+    (cloned_keys, cloned_values, _), (original_keys, original_values, _) = (
+        next(iter(cloned)),
+        next(iter(cache)),
+    )
+    assert torch.equal(cloned_keys, original_keys)
+    assert torch.equal(cloned_values, original_values)
+
+
+def _make_action_time_layers(action_dim=7, width=16, seed=0):
+    torch.manual_seed(seed)
+    return (
+        torch.nn.Linear(action_dim, width),
+        torch.nn.Linear(2 * width, width),
+        torch.nn.Linear(width, width),
+    )
+
+
+def test_fuse_action_time_embedding_matches_pi0_inline_loop():
+    # Verbatim pi0 block: time_emb cast to timestep.dtype, no gradient checkpointing.
+    in_proj, mlp_in, mlp_out = _make_action_time_layers()
+    min_period, max_period = 4e-3, 4.0
+    noisy_actions = torch.randn(3, 5, 7)
+    timestep = torch.rand(3)
+
+    time_emb = create_sinusoidal_pos_embedding(
+        timestep, in_proj.out_features, min_period=min_period, max_period=max_period, device=timestep.device
+    )
+    time_emb = time_emb.type(dtype=timestep.dtype)
+    action_emb = in_proj(noisy_actions)
+    time_emb = time_emb[:, None, :].expand_as(action_emb)
+    action_time_emb = torch.cat([action_emb, time_emb], dim=2)
+    ref = mlp_out(F.silu(mlp_in(action_time_emb)))
+
+    out = fuse_action_time_embedding(
+        noisy_actions,
+        timestep,
+        action_in_proj=in_proj,
+        action_time_mlp_in=mlp_in,
+        action_time_mlp_out=mlp_out,
+        min_period=min_period,
+        max_period=max_period,
+        time_emb_dtype=timestep.dtype,
+    )
+    assert torch.equal(out, ref)
+
+
+def test_fuse_action_time_embedding_matches_smolvla_inline_loop():
+    # Verbatim smolvla block: time_emb cast to action_emb.dtype (default), no checkpointing.
+    in_proj, mlp_in, mlp_out = _make_action_time_layers(seed=1)
+    min_period, max_period = 4e-3, 4.0
+    noisy_actions = torch.randn(2, 4, 7)
+    timestep = torch.rand(2)
+
+    action_emb = in_proj(noisy_actions)
+    dtype = action_emb.dtype
+    time_emb = create_sinusoidal_pos_embedding(
+        timestep, in_proj.out_features, min_period, max_period, device=action_emb.device
+    )
+    time_emb = time_emb.type(dtype=dtype)
+    time_emb = time_emb[:, None, :].expand_as(action_emb)
+    action_time_emb = torch.cat([action_emb, time_emb], dim=2)
+    action_time_emb = mlp_in(action_time_emb)
+    action_time_emb = F.silu(action_time_emb)
+    ref = mlp_out(action_time_emb)
+
+    out = fuse_action_time_embedding(
+        noisy_actions,
+        timestep,
+        action_in_proj=in_proj,
+        action_time_mlp_in=mlp_in,
+        action_time_mlp_out=mlp_out,
+        min_period=min_period,
+        max_period=max_period,
+    )
+    assert torch.equal(out, ref)
+
+
+def test_fuse_action_time_embedding_apply_hook_is_invoked():
+    # The apply hook (gradient-checkpoint wrapper in pi0/eo1) must wrap both sub-steps.
+    in_proj, mlp_in, mlp_out = _make_action_time_layers(seed=2)
+    calls = []
+
+    def apply(fn, arg):
+        calls.append(fn)
+        return fn(arg)
+
+    out = fuse_action_time_embedding(
+        torch.randn(2, 4, 7),
+        torch.rand(2),
+        action_in_proj=in_proj,
+        action_time_mlp_in=mlp_in,
+        action_time_mlp_out=mlp_out,
+        min_period=4e-3,
+        max_period=4.0,
+        apply=apply,
+    )
+    assert len(calls) == 2
+    assert out.shape == (2, 4, 16)

tests/policies/wall_x/test_wallx.py

@@ -25,13 +25,57 @@ pytest.importorskip("transformers")
 pytest.importorskip("torchdiffeq")
 
 from lerobot.policies.factory import make_policy_config  # noqa: E402
-from lerobot.policies.wall_x import WallXConfig  # noqa: E402
+from lerobot.policies.wall_x import (
+    WallXConfig,  # noqa: E402
+)
 from lerobot.policies.wall_x.modeling_wall_x import WallXPolicy  # noqa: E402
 from lerobot.policies.wall_x.processor_wall_x import make_wall_x_pre_post_processors  # noqa: E402
+from lerobot.policies.wall_x.qwen_model import Qwen2_5_VLMoEModel, Qwen2_5_VLTextConfig  # noqa: E402
 from lerobot.utils.random_utils import set_seed  # noqa: E402
 from tests.utils import require_cuda, require_hf_token  # noqa: E402
 
 
+def test_moe_model_captures_requested_hidden_states_and_attentions():
+    hidden_size = 16
+    expert_config = {
+        "hidden_size": hidden_size,
+        "intermediate_size": 32,
+        "hidden_act": "silu",
+    }
+    config = Qwen2_5_VLTextConfig(
+        vocab_size=32,
+        hidden_size=hidden_size,
+        intermediate_size=32,
+        num_hidden_layers=2,
+        num_attention_heads=4,
+        num_key_value_heads=4,
+        max_position_embeddings=32,
+        layer_types=["full_attention", "full_attention"],
+        rope_parameters={
+            "rope_type": "default",
+            "rope_theta": 1_000_000.0,
+            "mrope_section": [1, 1, 0],
+        },
+        num_experts=2,
+        experts=[expert_config, expert_config],
+        dim_inputs=(hidden_size, hidden_size),
+        mlp_moe=True,
+    )
+    config._attn_implementation = "eager"
+    model = Qwen2_5_VLMoEModel(config)
+    input_ids = torch.tensor([[1, 2, 3]])
+
+    output = model(
+        input_ids=input_ids,
+        moe_token_types=torch.zeros_like(input_ids),
+        output_hidden_states=True,
+        output_attentions=True,
+    )
+
+    assert len(output.hidden_states) == config.num_hidden_layers + 1
+    assert len(output.attentions) == config.num_hidden_layers
+
+
 @require_cuda
 @require_hf_token
 def test_policy_instantiation():