optmize topreward input processing (#3660)

docs/source/topreward.mdx

@@ -53,7 +53,7 @@ or, with `uv` from a source checkout:
 uv sync --extra topreward
 ```
 
-This pulls in `transformers` and `qwen-vl-utils`. The first time you run TOPReward, Hugging Face will also download the VLM weights from the Hub (~16 GB for Qwen3-VL-8B-Instruct). A GPU is strongly recommended.
+This pulls in `transformers`. The first time you run TOPReward, Hugging Face will also download the VLM weights from the Hub (~16 GB for Qwen3-VL-8B-Instruct). A GPU is strongly recommended.
 
 ## Model Inputs and Outputs
 

pyproject.toml

@@ -209,7 +209,7 @@ groot = [
     "flash-attn>=2.5.9,<3.0.0 ; sys_platform != 'darwin'"
 ]
 sarm = ["lerobot[transformers-dep]", "pydantic>=2.0.0,<3.0.0", "faker>=33.0.0,<35.0.0", "lerobot[matplotlib-dep]", "lerobot[qwen-vl-utils-dep]"]
-topreward = ["lerobot[transformers-dep]", "lerobot[qwen-vl-utils-dep]"]
+topreward = ["lerobot[transformers-dep]"]
 xvla = ["lerobot[transformers-dep]"]
 eo1 = ["lerobot[transformers-dep]", "lerobot[qwen-vl-utils-dep]"]
 hilserl = ["lerobot[transformers-dep]", "lerobot[dataset]", "gym-hil>=0.1.13,<0.2.0", "lerobot[grpcio-dep]", "lerobot[placo-dep]"]

src/lerobot/rewards/topreward/compute_rabc_weights.py

@@ -107,10 +107,10 @@ def compute_instruction_rewards_for_prefixes(
     else:
         prefix_lengths = np.unique(np.linspace(1, num_frames, num_samples).round().astype(np.int64))
 
+    episode_frames = torch.stack([dataset[ep_start + i][image_key] for i in range(num_frames)])
     rewards: list[float] = []
     for length in prefix_lengths:
-        frames = torch.stack([dataset[ep_start + i][image_key] for i in range(int(length))])
+        frames = episode_frames[: int(length)].unsqueeze(0)  # (1, T, C, H, W)
-        frames = frames.unsqueeze(0)  # (1, T, C, H, W)
 
         transition = {
             TransitionKey.OBSERVATION: {image_key: frames},
@@ -146,7 +146,6 @@ def compute_topreward_progress(
     device: str = "cuda",
     num_samples: int | None = None,
     fps: float | None = None,
-    reduction: str | None = None,
     episodes: list[int] | None = None,
 ) -> Path:
     """Run TOPReward over a dataset and write per-frame progress."""
@@ -154,10 +153,10 @@ def compute_topreward_progress(
         logging.info(f"Loading TOPReward config from: {reward_model_path}")
         model = TOPRewardModel.from_pretrained(reward_model_path)
         config = model.config
+        config.device = device
         if vlm_name is not None and vlm_name != config.vlm_name:
             logging.info(f"Overriding vlm_name from config: {config.vlm_name} -> {vlm_name}")
             config.vlm_name = vlm_name
-            config.device = device
             model = TOPRewardModel(config)
     else:
         config_kwargs: dict[str, Any] = {"device": device}
@@ -165,8 +164,6 @@ def compute_topreward_progress(
             config_kwargs["vlm_name"] = vlm_name
         if fps is not None:
             config_kwargs["fps"] = fps
-        if reduction is not None:
-            config_kwargs["reduction"] = reduction
         config = TOPRewardConfig(**config_kwargs)
         logging.info(f"Constructing TOPReward with VLM: {config.vlm_name}")
         model = TOPRewardModel(config)
@@ -302,9 +299,6 @@ Examples:
         help="Process only these episode indices (e.g. --episodes 0 or --episodes 0 5 10).",
     )
     parser.add_argument("--fps", type=float, default=None, help="Override TOPRewardConfig.fps.")
-    parser.add_argument(
-        "--reduction", type=str, default=None, choices=["mean", "sum"], help="Override reduction."
-    )
     parser.add_argument(
         "--push-to-hub", action="store_true", help="Upload to the dataset repo on HuggingFace Hub."
     )
@@ -321,7 +315,6 @@ Examples:
         device=args.device,
         num_samples=args.num_samples,
         fps=args.fps,
-        reduction=args.reduction,
         episodes=args.episodes,
     )
 

src/lerobot/rewards/topreward/configuration_topreward.py

@@ -22,8 +22,8 @@ from lerobot.utils.constants import OBS_IMAGES
 
 # Default prompt scaffolding from the upstream TOPReward paper / reference
 # implementation (``QwenClient.compute_instruction_reward``). The prompt
-# computes the log-likelihood of the suffix ``f"{instruction} ... True"``
+# scores the terminal ``True`` token in ``f"{instruction} ... True"``
-# given the video, then reduces those token log-probs to a scalar reward.
+# given the video.
 DEFAULT_PROMPT_PREFIX = (
     "The above video shows a robot manipulation trajectory that completes the following task: "
 )
@@ -67,8 +67,6 @@ class TOPRewardConfig(RewardModelConfig):
         add_chat_template: If ``True``, wrap the full prompt with the
             tokenizer's chat template before tokenisation (matches
             upstream ``add_chat_template=True``).
-        reduction: Reduction over per-token log-probs of the suffix
-            tokens (``"mean"`` or ``"sum"``).
         success_threshold: Optional log-prob threshold. If finite,
             :meth:`TOPRewardModel.compute_reward` returns
             ``(reward > success_threshold).float()`` instead of the raw
@@ -96,7 +94,6 @@ class TOPRewardConfig(RewardModelConfig):
     prompt_suffix_template: str = DEFAULT_PROMPT_SUFFIX_TEMPLATE
     add_chat_template: bool = False
 
-    reduction: str = "mean"
     success_threshold: float = float("-inf")
     max_input_length: int = 32768
 
@@ -116,8 +113,6 @@ class TOPRewardConfig(RewardModelConfig):
 
     def __post_init__(self) -> None:
         super().__post_init__()
-        if self.reduction not in {"mean", "sum"}:
-            raise ValueError(f"reduction must be 'mean' or 'sum', got {self.reduction!r}")
         if self.max_frames is not None and self.max_frames < 1:
             raise ValueError(f"max_frames must be >= 1, got {self.max_frames}")
         if self.fps <= 0:

src/lerobot/rewards/topreward/modeling_topreward.py

@@ -29,15 +29,16 @@ and returns that log-likelihood as the reward signal.
 Inference recipe:
 
 1. The processor builds a chat-style prompt, tokenises it, and emits
-   ``input_ids``, ``attention_mask``, vision tensors, and ``prompt_length``.
+   ``input_ids``, ``attention_mask``, vision tensors, and ``labels``.
-2. The model label-masks everything before ``prompt_length`` with ``-100``.
+   The processor label-masks everything except the terminal answer token with
-3. Forward the full token sequence through the VLM.
+   ``-100``.
-4. Read per-token log-probabilities of the unmasked suffix tokens from the
+2. Forward the full token sequence through the VLM.
-   logits and reduce them (mean or sum) into a scalar reward.
+3. Read the terminal answer token log-probability from the logits as the
+   scalar reward.
 
-With the default ``prompt_suffix_template`` and ``prompt_length = input_len - 1``
+With the default ``prompt_suffix_template``, the only unmasked token is the
-(mirrored from upstream), the only unmasked token is the literal ``"True"``
+literal ``"True"`` at the end — the reward is
-at the end — the reward is ``log P("True" | video + prompt + instruction)``.
+``log P("True" | video + prompt + instruction)``.
 
 This LeRobot port is **inference-only and not trainable** — :meth:`forward`
 is intentionally inherited from :class:`PreTrainedRewardModel` and raises
@@ -66,6 +67,7 @@ from huggingface_hub import HfApi, hf_hub_download
 from huggingface_hub.constants import CONFIG_NAME
 from huggingface_hub.errors import HfHubHTTPError
 from torch import Tensor
+from torch.nn.functional import cross_entropy
 
 from lerobot.configs.rewards import RewardModelConfig
 from lerobot.rewards.pretrained import PreTrainedRewardModel
@@ -116,57 +118,29 @@ class TOPRewardModel(PreTrainedRewardModel):
 
     def compute_reward(self, batch: dict[str, Any]) -> Tensor:
         """Return one log-prob reward per sample in the batch."""
-        inputs = {
+        inputs: dict[str, Any] = {}
-            key: batch[f"{TOPREWARD_FEATURE_PREFIX}{key}"]
+        for key in TOPREWARD_INPUT_KEYS:
-            for key in TOPREWARD_INPUT_KEYS
+            batch_key = f"{TOPREWARD_FEATURE_PREFIX}{key}"
-            if f"{TOPREWARD_FEATURE_PREFIX}{key}" in batch
+            if batch_key not in batch:
-        }
-        if "input_ids" not in inputs:
                 raise KeyError(
-                f"TOPReward batch missing pre-encoded inputs (expected "
+                    f"TOPReward batch missing `{batch_key}`. Make sure the "
-                f"`{TOPREWARD_FEATURE_PREFIX}input_ids`). Make sure the "
                     "TOPRewardEncoderProcessorStep ran before `compute_reward`."
                 )
+            inputs[key] = batch[batch_key]
 
-        prompt_lengths = inputs.pop("prompt_length")
         device = next(self.model.parameters()).device
         inputs = {key: value.to(device) if hasattr(value, "to") else value for key, value in inputs.items()}
-
+        labels = inputs.pop("labels")
-        labels = inputs["input_ids"].clone()
+        inputs["logits_to_keep"] = 2
-        for i, plen in enumerate(prompt_lengths.tolist()):
-            labels[i, : int(plen)] = -100
-        if "attention_mask" in inputs:
-            labels = labels.masked_fill(inputs["attention_mask"] == 0, -100)
 
         self.eval()
         with torch.no_grad():
             outputs = self.model(**inputs)
-
+        logits = outputs.logits
-        logits = outputs.logits[:, :-1, :]
+        rewards = -cross_entropy(logits[:, -2, :].float(), labels[:, -1], reduction="none")
-        target_labels = labels[:, 1:]
-        log_probs = torch.nn.functional.log_softmax(logits, dim=-1)
-        mask = target_labels != -100
-        safe_targets = target_labels.masked_fill(~mask, 0)
-        token_log_probs = log_probs.gather(-1, safe_targets.unsqueeze(-1)).squeeze(-1)
-
-        batch_size = inputs["input_ids"].shape[0]
-        rewards = []
-        for i in range(batch_size):
-            sample_log_probs = token_log_probs[i][mask[i]]
-            if sample_log_probs.numel() == 0:
-                raise RuntimeError(
-                    "TOPReward could not isolate any suffix tokens to score. Check that "
-                    "`prompt_suffix_template` produces at least one tokenised character."
-                )
-            if self.config.reduction == "sum":
-                rewards.append(sample_log_probs.sum().item())
-            else:
-                rewards.append(sample_log_probs.mean().item())
-
-        out = torch.as_tensor(rewards, dtype=torch.float32)
         if np.isfinite(self.config.success_threshold):
-            out = (out > self.config.success_threshold).float()
+            rewards = (rewards > self.config.success_threshold).float()
-        return out.to(self.config.device or "cpu")
+        return rewards.to(self.config.device or "cpu")
 
     def _save_pretrained(self, save_directory: Path) -> None:
         """Save ``config.json`` only."""

src/lerobot/rewards/topreward/processor_topreward.py

@@ -19,9 +19,7 @@ from __future__ import annotations
 from dataclasses import dataclass, field
 from typing import TYPE_CHECKING, Any
 
-import numpy as np
 import torch
-from PIL import Image
 from torch import Tensor
 
 from lerobot.configs import PipelineFeatureType, PolicyFeature
@@ -60,39 +58,33 @@ _TRUE_ANSWER = "True"
 TOPREWARD_VLM_INPUT_KEYS = (
     "input_ids",
     "attention_mask",
-    "pixel_values",
     "pixel_values_videos",
-    "image_grid_thw",
     "video_grid_thw",
-    "second_per_grid_ts",
     "mm_token_type_ids",
 )
-TOPREWARD_METADATA_KEYS = ("prompt_length",)
+TOPREWARD_INPUT_KEYS = TOPREWARD_VLM_INPUT_KEYS + ("labels",)
-TOPREWARD_INPUT_KEYS = TOPREWARD_VLM_INPUT_KEYS + TOPREWARD_METADATA_KEYS
 
 
-def _video_to_numpy(video: Tensor, *, max_frames: int | None) -> np.ndarray:
+def _prepare_video_batch(video: Tensor, *, max_frames: int | None) -> Tensor:
-    """Convert one trajectory tensor to a ``(T, H, W, C) uint8`` numpy array."""
+    """Return videos as ``(B, T, C, H, W)`` uint8 tensors for Qwen3-VL."""
+    if video.ndim == 4:
+        video = video.unsqueeze(1)
+    elif video.ndim != 5:
+        raise ValueError(
+            f"Expected TOPReward frames with shape (B,C,H,W) or (B,T,C,H,W); got {tuple(video.shape)}"
+        )
+
     if max_frames is not None:
-        video = video[-max_frames:]
+        video = video[:, -max_frames:]
-    if video.shape[1] in (1, 3):
+    if video.shape[-1] in (1, 3):
-        video = video.permute(0, 2, 3, 1)
+        video = video.permute(0, 1, 4, 2, 3)
-    elif video.shape[-1] not in (1, 3):
+    elif video.shape[2] not in (1, 3):
         raise ValueError(f"Expected channel dim of size 1 or 3, got shape {tuple(video.shape)}")
 
-    array = video.detach().cpu().numpy()
+    if video.is_floating_point():
-    if np.issubdtype(array.dtype, np.floating) and array.size > 0 and array.max() <= 1.0:
+        video = video * 255.0
-        array = array * 255.0
-    return np.clip(array, 0, 255).astype(np.uint8)
 
-
+    return video.clamp(0, 255).to(torch.uint8).contiguous()
-def _frames_to_pil(frames: np.ndarray) -> list[Image.Image]:
-    """Convert ``(T, H, W, C)`` uint8 frames to a list of PIL images."""
-    if frames.ndim != 4:
-        raise ValueError(f"Expected (T,H,W,C) frames; got shape {frames.shape}")
-    if frames.dtype != np.uint8:
-        frames = np.clip(frames, 0, 255).astype(np.uint8)
-    return [Image.fromarray(frames[i]) for i in range(frames.shape[0])]
 
 
 def _expand_tasks(task: Any, *, batch_size: int, default: str | None) -> list[str]:
@@ -120,10 +112,9 @@ class TOPRewardEncoderProcessorStep(ProcessorStep):
 
     Loads a :class:`~transformers.AutoProcessor` matching ``vlm_name`` and
     builds the full chat prompt including the instruction suffix. The
-    resulting ``input_ids``, ``attention_mask``, vision tensors, and a
+    resulting ``input_ids``, ``attention_mask``, vision tensors, and
-    per-sample ``prompt_length`` integer are written under the
+    ``labels`` are written under the ``observation.topreward.*`` namespace
-    ``observation.topreward.*`` namespace so the model can label-mask and
+    so the model can score without re-tokenising.
-    forward without re-tokenising.
 
     At call time the step reads:
 
@@ -131,7 +122,7 @@ class TOPRewardEncoderProcessorStep(ProcessorStep):
     - ``complementary_data[task_key]``: a string or list of strings.
 
     and writes ``observation[f"{TOPREWARD_FEATURE_PREFIX}<name>"]`` for the
-    Qwen-VL tensors plus ``prompt_length``.
+    Qwen-VL tensors plus ``labels``.
     """
 
     vlm_name: str = "Qwen/Qwen3-VL-8B-Instruct"
@@ -149,35 +140,26 @@ class TOPRewardEncoderProcessorStep(ProcessorStep):
 
     def __post_init__(self) -> None:
         require_package("transformers", extra="topreward")
-        require_package("qwen-vl-utils", extra="topreward", import_name="qwen_vl_utils")
         self._processor = AutoProcessor.from_pretrained(self.vlm_name, trust_remote_code=True)
 
     def __call__(self, transition: EnvTransition) -> EnvTransition:
         observation = transition.get(TransitionKey.OBSERVATION)
         complementary = transition.get(TransitionKey.COMPLEMENTARY_DATA) or {}
-        if not isinstance(observation, dict):
-            raise ValueError("TOPRewardEncoderProcessorStep requires an observation dict")
-
         if self.image_key not in observation:
             raise KeyError(f"TOPReward expected image key {self.image_key!r} in observation")
 
         frames = observation[self.image_key]
-        tensor = frames.detach().cpu() if isinstance(frames, Tensor) else torch.as_tensor(frames)
+        videos = frames.detach().cpu() if isinstance(frames, Tensor) else torch.as_tensor(frames)
-        if tensor.ndim == 4:
+        videos = _prepare_video_batch(videos, max_frames=self.max_frames)
-            tensor = tensor.unsqueeze(1)
-        elif tensor.ndim != 5:
-            raise ValueError(
-                f"Expected TOPReward frames with shape (B,C,H,W) or (B,T,C,H,W); got {tuple(tensor.shape)}"
-            )
 
-        batch_size = tensor.shape[0]
+        batch_size = videos.shape[0]
         tasks = _expand_tasks(
             complementary.get(self.task_key, self.default_task),
             batch_size=batch_size,
             default=self.default_task,
         )
 
-        encoded = self._encode_batch(tensor, tasks)
+        encoded = self._encode_batch(videos, tasks, batch_size)
 
         new_observation = dict(observation)
         for key, value in encoded.items():
@@ -187,34 +169,33 @@ class TOPRewardEncoderProcessorStep(ProcessorStep):
         new_transition[TransitionKey.OBSERVATION] = new_observation
         return new_transition
 
-    def _encode_batch(self, tensor: Tensor, tasks: list[str]) -> dict[str, Any]:
+    def _encode_batch(self, videos: Tensor, tasks: list[str], batch_size) -> dict[str, Any]:
         """Tokenise a batch of (frames, task) pairs into Qwen-VL tensors.
 
-        Processes samples one at a time (each may have a different token
+        The loop only builds per-sample chat strings. Tokenisation, padding,
-        length due to different numbers of vision patches), then pads /
+        video preprocessing, and label construction are batched.
-        stacks the results.
         """
-        from qwen_vl_utils import process_vision_info
 
-        batch_size = tensor.shape[0]
+        texts: list[str] = []
-        all_encoded: list[dict[str, Any]] = []
+        video_metadata = [
-        all_prompt_lengths: list[int] = []
+            {
-
+                "total_num_frames": int(videos.shape[1]),
-        for i in range(batch_size):
+                "fps": float(self.fps),
-            frames_np = _video_to_numpy(tensor[i], max_frames=self.max_frames)
+                "frames_indices": list(range(int(videos.shape[1]))),
-            pil_frames = _frames_to_pil(frames_np)
+            }
-            task = tasks[i]
+            for _ in range(batch_size)
-
+        ]
-            instruction_suffix = self.prompt_suffix_template.format(instruction=task)
         eos_token = self._processor.tokenizer.eos_token
 
+        for i in range(batch_size):
+            instruction_suffix = self.prompt_suffix_template.format(instruction=tasks[i])
             if self.add_chat_template:
                 suffix_for_template = instruction_suffix.removesuffix(_TRUE_ANSWER).rstrip()
                 templated_messages = [
                     {
                         "role": "user",
                         "content": [
-                            {"type": "video", "video": pil_frames, "fps": self.fps},
+                            {"type": "video", "video": videos[i], "fps": self.fps},
                             {"type": "text", "text": f"{self.prompt_prefix}{suffix_for_template}"},
                         ],
                     }
@@ -223,13 +204,12 @@ class TOPRewardEncoderProcessorStep(ProcessorStep):
                     templated_messages, tokenize=False, add_generation_prompt=True
                 )
                 full_text = f"{prompt_chat}{_TRUE_ANSWER}"
-                image_inputs, video_inputs = process_vision_info(templated_messages)
             else:
                 user_messages = [
                     {
                         "role": "user",
                         "content": [
-                            {"type": "video", "video": pil_frames, "fps": self.fps},
+                            {"type": "video", "video": videos[i], "fps": self.fps},
                             {"type": "text", "text": self.prompt_prefix},
                         ],
                     }
@@ -240,59 +220,29 @@ class TOPRewardEncoderProcessorStep(ProcessorStep):
                 if eos_token is not None:
                     prompt_chat = prompt_chat.split(eos_token)[0]
                 full_text = f"{prompt_chat}{instruction_suffix}"
-                image_inputs, video_inputs = process_vision_info(user_messages)
 
-            inputs = self._processor(
+            texts.append(full_text)
-                text=[full_text],
+
-                images=image_inputs,
+        result = self._processor(
-                videos=video_inputs,
+            text=texts,
+            videos=videos,
+            video_metadata=video_metadata,
+            do_sample_frames=False,
             padding=True,
+            padding_side="left",
             return_tensors="pt",
         )
+        input_ids = result["input_ids"]
 
-            input_len = int(inputs["input_ids"].shape[-1])
+        if input_ids.shape[-1] > self.max_length:
-            if input_len > self.max_length:
             raise ValueError(
-                    f"TOPReward input length {input_len} exceeds max_length "
+                f"TOPReward input length {input_ids.shape[-1]} exceeds max_length "
                 f"{self.max_length}; lower `max_frames` or raise `max_length`."
             )
 
-            prompt_length = input_len - 1
+        labels = torch.full_like(input_ids, -100)
-            all_encoded.append(inputs)
+        labels[:, -1] = input_ids[:, -1]
-            all_prompt_lengths.append(prompt_length)
+        result["labels"] = labels
-
-        result = dict(all_encoded[0]) if batch_size == 1 else self._pad_and_stack(all_encoded)
-
-        result["prompt_length"] = torch.tensor(all_prompt_lengths, dtype=torch.long)
-        return result
-
-    @staticmethod
-    def _pad_and_stack(encoded_list: list[dict[str, Any]]) -> dict[str, Any]:
-        """Right-pad and stack per-sample encoded dicts into a batch."""
-        keys = [k for k in encoded_list[0] if isinstance(encoded_list[0][k], Tensor)]
-        max_len = max(enc["input_ids"].shape[-1] for enc in encoded_list)
-        result: dict[str, Any] = {}
-
-        for key in keys:
-            tensors = [enc[key] for enc in encoded_list]
-            if key in ("input_ids", "attention_mask"):
-                padded = []
-                pad_value = 0
-                for t in tensors:
-                    pad_size = max_len - t.shape[-1]
-                    if pad_size > 0:
-                        padded.append(torch.nn.functional.pad(t, (0, pad_size), value=pad_value))
-                    else:
-                        padded.append(t)
-                result[key] = torch.cat(padded, dim=0)
-            else:
-                # Vision tensors (pixel_values_videos, image_grid_thw, etc.) are expected
-                # to have matching shapes since max_frames is applied uniformly per batch
-                result[key] = torch.cat(tensors, dim=0)
-
-        for key in encoded_list[0]:
-            if key not in result:
-                result[key] = encoded_list[0][key]
         return result
 
     def transform_features(
@@ -325,8 +275,8 @@ def make_topreward_pre_post_processors(
     """Pipeline that pre-encodes frames + task into Qwen-VL tensors.
 
     The preprocessor adds a batch dimension if needed, runs TOPReward's
-    encoder (which tokenises the full prompt and emits ``prompt_length``),
+    encoder (which tokenises the full prompt and emits ``labels``), and
-    and moves everything to the configured device. The postprocessor is
+    moves everything to the configured device. The postprocessor is
     the identity since TOPReward outputs a single reward tensor.
     """
     preprocessor = PolicyProcessorPipeline[dict[str, Any], dict[str, Any]](

tests/rewards/test_modeling_topreward.py

@@ -24,7 +24,7 @@ import torch
 from lerobot.configs.rewards import RewardModelConfig
 from lerobot.rewards.factory import get_reward_model_class, make_reward_model_config
 from lerobot.rewards.topreward import TOPRewardConfig
-from lerobot.rewards.topreward.processor_topreward import TOPREWARD_FEATURE_PREFIX
+from lerobot.rewards.topreward.processor_topreward import TOPREWARD_FEATURE_PREFIX, TOPREWARD_INPUT_KEYS
 from tests.utils import skip_if_package_missing
 
 
@@ -45,20 +45,23 @@ class _FakeQwenModel(torch.nn.Module):
     def from_pretrained(cls, *args, **kwargs):  # noqa: ARG003
         return cls()
 
-    def forward(self, input_ids, attention_mask=None, labels=None, **kwargs):  # noqa: ARG002
+    def forward(  # noqa: ARG002
+        self, input_ids, attention_mask=None, labels=None, logits_to_keep=0, **kwargs
+    ):
         batch_size, seq_len = input_ids.shape
         vocab_size = 1000
         logits = torch.zeros(batch_size, seq_len, vocab_size)
         # Place a controlled log-prob at the target token position so the
         # model returns a predictable reward value.
-        # The label-masked suffix is the last token (prompt_length = seq_len - 1).
+        # The label-masked suffix is the last token.
         # After the causal-LM shift (logits[:, :-1], labels[:, 1:]) the scored
         # position is logits[:, -2, :] predicting labels[:, -1].
         # We set logits so that log_softmax at the target token ≈ _reward_value.
-        if labels is not None:
         for i in range(batch_size):
             target_idx = int(input_ids[i, -1].item())
             logits[i, -2, target_idx] = self._reward_value * -10  # high logit -> high log-prob
+        if logits_to_keep:
+            logits = logits[:, -logits_to_keep:, :]
         return SimpleNamespace(logits=logits)
 
 
@@ -72,17 +75,39 @@ def _patch_build(monkeypatch) -> None:
 def _make_batch(
     input_ids: torch.Tensor,
     attention_mask: torch.Tensor | None = None,
-    prompt_length: torch.Tensor | None = None,
+    labels: torch.Tensor | None = None,
+    *,
+    omit: str | None = None,
 ) -> dict[str, torch.Tensor]:
     """Build a ``compute_reward``-ready batch using TOPReward's namespaced keys."""
-    batch: dict[str, torch.Tensor] = {f"{TOPREWARD_FEATURE_PREFIX}input_ids": input_ids}
+    batch_size, seq_len = input_ids.shape
-    if attention_mask is not None:
+    if attention_mask is None:
-        batch[f"{TOPREWARD_FEATURE_PREFIX}attention_mask"] = attention_mask
+        attention_mask = torch.ones(batch_size, seq_len, dtype=torch.long)
-    if prompt_length is not None:
+    batch: dict[str, torch.Tensor] = {}
-        batch[f"{TOPREWARD_FEATURE_PREFIX}prompt_length"] = prompt_length
+    if labels is not None:
+        batch[f"{TOPREWARD_FEATURE_PREFIX}labels"] = labels
+    batch.update(
+        {
+            f"{TOPREWARD_FEATURE_PREFIX}input_ids": input_ids,
+            f"{TOPREWARD_FEATURE_PREFIX}attention_mask": attention_mask,
+            f"{TOPREWARD_FEATURE_PREFIX}pixel_values_videos": torch.zeros(
+                batch_size, 1536, dtype=torch.float32
+            ),
+            f"{TOPREWARD_FEATURE_PREFIX}video_grid_thw": torch.ones(batch_size, 3, dtype=torch.long),
+            f"{TOPREWARD_FEATURE_PREFIX}mm_token_type_ids": torch.zeros_like(input_ids),
+        }
+    )
+    if omit is not None:
+        batch.pop(f"{TOPREWARD_FEATURE_PREFIX}{omit}", None)
     return batch
 
 
+def _terminal_labels(input_ids: torch.Tensor) -> torch.Tensor:
+    labels = torch.full_like(input_ids, -100)
+    labels[:, -1] = input_ids[:, -1]
+    return labels
+
+
 # ---------------------------------------------------------------------------
 # Registry + factory
 # ---------------------------------------------------------------------------
@@ -105,11 +130,6 @@ def test_topreward_factory_returns_in_tree_class():
 # ---------------------------------------------------------------------------
 
 
-def test_topreward_config_rejects_bad_reduction():
-    with pytest.raises(ValueError, match="reduction must be"):
-        TOPRewardConfig(device="cpu", reduction="median")
-
-
 def test_topreward_config_rejects_zero_max_frames():
     with pytest.raises(ValueError, match="max_frames must be >= 1"):
         TOPRewardConfig(device="cpu", max_frames=0)
@@ -142,9 +162,9 @@ def test_topreward_compute_reward_returns_one_scalar_per_sample(monkeypatch):
 
     input_ids = torch.randint(0, 100, (2, 10))
     attention_mask = torch.ones(2, 10, dtype=torch.long)
-    prompt_length = torch.tensor([9, 9])  # unmask only the last token
+    labels = _terminal_labels(input_ids)
 
-    batch = _make_batch(input_ids, attention_mask, prompt_length)
+    batch = _make_batch(input_ids, attention_mask, labels)
     rewards = model.compute_reward(batch)
 
     assert rewards.shape == (2,)
@@ -162,9 +182,9 @@ def test_topreward_compute_reward_applies_success_threshold(monkeypatch):
 
     input_ids = torch.randint(0, 100, (2, 10))
     attention_mask = torch.ones(2, 10, dtype=torch.long)
-    prompt_length = torch.tensor([9, 9])
+    labels = _terminal_labels(input_ids)
 
-    batch = _make_batch(input_ids, attention_mask, prompt_length)
+    batch = _make_batch(input_ids, attention_mask, labels)
     rewards = model.compute_reward(batch)
 
     assert rewards.shape == (2,)
@@ -180,7 +200,37 @@ def test_topreward_compute_reward_errors_when_inputs_missing(monkeypatch):
     model = TOPRewardModel(cfg)
 
     with pytest.raises(KeyError, match=r"observation\.topreward\.input_ids"):
-        model.compute_reward({})
+        model.compute_reward(_make_batch(torch.randint(0, 100, (1, 10)), omit="input_ids"))
+
+
+@skip_if_package_missing("transformers")
+def test_topreward_compute_reward_errors_when_labels_missing(monkeypatch):
+    from lerobot.rewards.topreward.modeling_topreward import TOPRewardModel
+
+    _patch_build(monkeypatch)
+    cfg = TOPRewardConfig(device="cpu")
+    model = TOPRewardModel(cfg)
+
+    input_ids = torch.randint(0, 100, (1, 10))
+    with pytest.raises(KeyError, match=r"observation\.topreward\.labels"):
+        model.compute_reward(_make_batch(input_ids, labels=None))
+
+
+@skip_if_package_missing("transformers")
+def test_topreward_compute_reward_requires_all_encoder_keys(monkeypatch):
+    from lerobot.rewards.topreward.modeling_topreward import TOPRewardModel
+
+    _patch_build(monkeypatch)
+    cfg = TOPRewardConfig(device="cpu")
+    model = TOPRewardModel(cfg)
+
+    input_ids = torch.randint(0, 100, (1, 10))
+    labels = _terminal_labels(input_ids)
+    required_encoder_keys = set(TOPREWARD_INPUT_KEYS) - {"input_ids", "labels"}
+
+    for key in required_encoder_keys:
+        with pytest.raises(KeyError, match=rf"observation\.topreward\.{key}"):
+            model.compute_reward(_make_batch(input_ids, labels=labels, omit=key))
 
 
 # ---------------------------------------------------------------------------
@@ -198,7 +248,6 @@ def test_topreward_save_pretrained_writes_only_config_json(monkeypatch, tmp_path
     cfg = TOPRewardConfig(
         device="cpu",
         vlm_name="Qwen/Qwen3-VL-8B-Instruct",
-        reduction="sum",
         fps=4.0,
         image_key="observation.images.front",
     )
@@ -217,7 +266,6 @@ def test_topreward_from_pretrained_local_dir_roundtrips_config(monkeypatch, tmp_
     cfg = TOPRewardConfig(
         device="cpu",
         vlm_name="Qwen/Qwen3-VL-8B-Instruct",
-        reduction="sum",
         fps=4.0,
         image_key="observation.images.front",
         add_chat_template=True,
@@ -229,7 +277,6 @@ def test_topreward_from_pretrained_local_dir_roundtrips_config(monkeypatch, tmp_
 
     assert isinstance(reloaded.config, TOPRewardConfig)
     assert reloaded.config.vlm_name == "Qwen/Qwen3-VL-8B-Instruct"
-    assert reloaded.config.reduction == "sum"
     assert reloaded.config.fps == 4.0
     assert reloaded.config.image_key == "observation.images.front"
     assert reloaded.config.add_chat_template is True

tests/rewards/test_topreward.py

@@ -0,0 +1,80 @@
+# 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.
+
+"""End-to-end TOPReward smoke test with the real Qwen3-VL model."""
+
+import os
+
+import pytest
+import torch
+
+pytest.importorskip("transformers")
+
+from lerobot.rewards.topreward.configuration_topreward import TOPRewardConfig  # noqa: E402
+from lerobot.rewards.topreward.modeling_topreward import TOPRewardModel  # noqa: E402
+from lerobot.rewards.topreward.processor_topreward import (  # noqa: E402
+    TOPREWARD_FEATURE_PREFIX,
+    TOPREWARD_INPUT_KEYS,
+    make_topreward_pre_post_processors,
+)
+from tests.utils import require_cuda  # noqa: E402
+
+pytestmark = pytest.mark.skipif(
+    os.environ.get("CI") == "true" or os.environ.get("GITHUB_ACTIONS") == "true",
+    reason="This test requires downloading and loading Qwen3-VL and is not meant for CI",
+)
+
+
+def _make_dummy_topreward_batch(image_key: str, task_key: str) -> dict[str, object]:
+    num_frames = 4
+    image_size = 64
+    frames = torch.zeros(1, num_frames, 3, image_size, image_size, dtype=torch.uint8)
+    for frame_idx in range(num_frames):
+        frames[0, frame_idx, 0].fill_(min(frame_idx * 48, 255))
+        frames[0, frame_idx, 1].fill_(96)
+        frames[0, frame_idx, 2].fill_(192)
+
+    return {
+        image_key: frames,
+        task_key: ["pick up the red cube"],
+    }
+
+
+@require_cuda
+def test_topreward_full_qwen3vl_preprocessor_to_compute_reward():
+    cfg = TOPRewardConfig(
+        vlm_name="Qwen/Qwen3-VL-8B-Instruct",
+        device="cuda",
+        max_frames=4,
+        fps=2.0,
+        max_input_length=4096,
+    )
+
+    preprocessor, _ = make_topreward_pre_post_processors(cfg)
+    encoded_batch = preprocessor(_make_dummy_topreward_batch(cfg.image_key, cfg.task_key))
+    for key in TOPREWARD_INPUT_KEYS:
+        assert f"{TOPREWARD_FEATURE_PREFIX}{key}" in encoded_batch
+
+    model = TOPRewardModel(cfg)
+    try:
+        model.to(cfg.device)
+        model.eval()
+        rewards = model.compute_reward(encoded_batch)
+    finally:
+        del model
+        torch.cuda.empty_cache()
+
+    assert rewards.shape == (1,)
+    assert rewards.dtype == torch.float32
+    assert torch.isfinite(rewards).all()

tests/rewards/test_topreward_processor.py

@@ -16,71 +16,71 @@
 
 from __future__ import annotations
 
-import numpy as np
 import pytest
 import torch
 
 from lerobot.configs import FeatureType, PipelineFeatureType, PolicyFeature
 from lerobot.rewards.topreward.processor_topreward import (
     TOPREWARD_FEATURE_PREFIX,
+    TOPREWARD_INPUT_KEYS,
     _expand_tasks,
-    _video_to_numpy,
+    _prepare_video_batch,
 )
 from lerobot.types import TransitionKey
 from tests.utils import skip_if_package_missing
 
 # ---------------------------------------------------------------------------
-# _video_to_numpy — pure (T, C, H, W) -> (T, H, W, C) uint8 conversion
+# _prepare_video_batch — raw image/video batch -> (B, T, C, H, W) uint8
 # ---------------------------------------------------------------------------
 
 
-def test_video_to_numpy_chw_float_is_converted_to_thwc_uint8():
+def test_prepare_video_batch_batched_chw_float_is_converted_to_uint8():
-    video = torch.rand(4, 3, 8, 8)
+    video = torch.rand(2, 4, 3, 8, 8)
-    array = _video_to_numpy(video, max_frames=None)
+    tensor = _prepare_video_batch(video, max_frames=None)
 
-    assert array.shape == (4, 8, 8, 3)
+    assert tensor.shape == (2, 4, 3, 8, 8)
-    assert array.dtype == np.uint8
+    assert tensor.dtype == torch.uint8
-    assert array.min() >= 0 and array.max() <= 255
+    assert tensor.min() >= 0 and tensor.max() <= 255
 
 
-def test_video_to_numpy_already_thwc_uint8_passes_through():
+def test_prepare_video_batch_batched_thwc_uint8_is_permuted_to_channel_first():
-    video = torch.randint(0, 256, (3, 8, 8, 3), dtype=torch.uint8)
+    video = torch.randint(0, 256, (2, 3, 8, 8, 3), dtype=torch.uint8)
-    array = _video_to_numpy(video, max_frames=None)
+    tensor = _prepare_video_batch(video, max_frames=None)
 
-    assert array.shape == (3, 8, 8, 3)
+    assert tensor.shape == (2, 3, 3, 8, 8)
-    assert array.dtype == np.uint8
+    assert tensor.dtype == torch.uint8
 
 
-def test_video_to_numpy_max_frames_tail_crops_recent_frames():
+def test_prepare_video_batch_max_frames_tail_crops_recent_frames():
-    video = torch.zeros(10, 3, 4, 4)
+    video = torch.zeros(1, 10, 3, 4, 4)
     for t in range(10):
-        video[t] = t / 9.0
+        video[:, t] = t / 9.0
 
-    array = _video_to_numpy(video, max_frames=3)
+    tensor = _prepare_video_batch(video, max_frames=3)
 
-    assert array.shape == (3, 4, 4, 3)
+    assert tensor.shape == (1, 3, 3, 4, 4)
-    assert int(array[0, 0, 0, 0]) == int(round(7 / 9 * 255))
+    assert int(tensor[0, 0, 0, 0, 0]) == int(7 / 9 * 255)
-    assert int(array[-1, 0, 0, 0]) == 255
+    assert int(tensor[0, -1, 0, 0, 0]) == 255
 
 
-def test_video_to_numpy_rejects_3d_input():
+def test_prepare_video_batch_rejects_3d_input():
-    with pytest.raises(ValueError, match="Expected channel dim"):
+    with pytest.raises(ValueError, match="Expected TOPReward frames"):
-        _video_to_numpy(torch.zeros(4, 8, 8), max_frames=None)
+        _prepare_video_batch(torch.zeros(4, 8, 8), max_frames=None)
 
 
-def test_video_to_numpy_floats_above_one_pass_through_without_rescaling():
+def test_prepare_video_batch_floats_above_one_are_rescaled_and_clipped():
-    video = torch.full((1, 3, 2, 2), 5.0)
+    video = torch.full((1, 1, 3, 2, 2), 5.0)
-    array = _video_to_numpy(video, max_frames=None)
+    tensor = _prepare_video_batch(video, max_frames=None)
 
-    assert array.shape == (1, 2, 2, 3)
+    assert tensor.shape == (1, 1, 3, 2, 2)
-    assert int(array.max()) == 5
+    assert int(tensor.max()) == 255
 
 
-def test_video_to_numpy_clips_very_large_floats_to_uint8_max():
+def test_prepare_video_batch_clips_very_large_floats_to_uint8_max():
-    video = torch.full((1, 3, 2, 2), 300.0)
+    video = torch.full((1, 1, 3, 2, 2), 300.0)
-    array = _video_to_numpy(video, max_frames=None)
+    tensor = _prepare_video_batch(video, max_frames=None)
 
-    assert int(array.max()) == 255
+    assert int(tensor.max()) == 255
 
 
 # ---------------------------------------------------------------------------
@@ -124,12 +124,11 @@ def test_expand_tasks_wrong_type_raises():
 
 
 # ---------------------------------------------------------------------------
-# Encoder step — stubbed AutoProcessor + process_vision_info
+# Encoder step — stubbed AutoProcessor
 # ---------------------------------------------------------------------------
 
 
 def _skip_if_topreward_extras_missing(func):
-    func = skip_if_package_missing("qwen-vl-utils", import_name="qwen_vl_utils")(func)
     func = skip_if_package_missing("transformers")(func)
     return func
 
@@ -155,32 +154,20 @@ class _FakeAutoProcessor:
 
     def __call__(self, text=None, images=None, videos=None, **kwargs):  # noqa: ARG002
         seq_len = 10
+        batch_size = len(text) if isinstance(text, list) else 1
         return {
-            "input_ids": torch.randint(0, 100, (1, seq_len)),
+            "input_ids": torch.randint(0, 100, (batch_size, seq_len)),
-            "attention_mask": torch.ones(1, seq_len, dtype=torch.long),
+            "attention_mask": torch.ones(batch_size, seq_len, dtype=torch.long),
+            "pixel_values_videos": torch.zeros(batch_size, 1536, dtype=torch.float32),
+            "video_grid_thw": torch.ones(batch_size, 3, dtype=torch.long),
+            "mm_token_type_ids": torch.zeros(batch_size, seq_len, dtype=torch.long),
         }
 
 
 def _build_step(monkeypatch, **overrides):
-    import importlib
-    import sys
-    import types
-
     from lerobot.rewards.topreward import processor_topreward
-    from lerobot.utils import import_utils
 
     monkeypatch.setattr(processor_topreward, "AutoProcessor", _FakeAutoProcessor)
-
-    # Stub qwen_vl_utils as a real module object (not MagicMock) so
-    # ``require_package`` / ``find_spec`` don't choke on a missing ``__spec__``.
-    fake_qwen_vl = types.ModuleType("qwen_vl_utils")
-    fake_qwen_vl.process_vision_info = lambda messages: (None, None)  # type: ignore[attr-defined]
-    fake_qwen_vl.__spec__ = importlib.machinery.ModuleSpec("qwen_vl_utils", None)
-    monkeypatch.setitem(sys.modules, "qwen_vl_utils", fake_qwen_vl)
-
-    # Clear the require_package cache so the stub is picked up.
-    import_utils._require_package_cache.pop("qwen_vl_utils", None)
-
     return processor_topreward.TOPRewardEncoderProcessorStep(**overrides)
 
 
@@ -192,27 +179,29 @@ def _make_transition(observation: dict, complementary: dict | None = None) -> di
 
 
 @_skip_if_topreward_extras_missing
-def test_encoder_step_emits_input_ids_and_prompt_length(monkeypatch):
+def test_encoder_step_emits_input_ids_and_labels(monkeypatch):
     """The processor must emit Qwen-VL tensors including ``input_ids`` and
-    ``prompt_length`` under the ``observation.topreward.*`` namespace."""
+    ``labels`` under the ``observation.topreward.*`` namespace."""
     step = _build_step(monkeypatch)
 
-    frames_batch = torch.zeros(1, 4, 3, 8, 8)
+    frames_batch = torch.zeros(2, 4, 3, 8, 8)
     out = step(
         _make_transition(
             observation={"observation.images.top": frames_batch},
-            complementary={"task": "pick"},
+            complementary={"task": ["pick", "place"]},
         )
     )
 
     obs_out = out[TransitionKey.OBSERVATION]
-    assert f"{TOPREWARD_FEATURE_PREFIX}input_ids" in obs_out
+    for key in TOPREWARD_INPUT_KEYS:
-    assert f"{TOPREWARD_FEATURE_PREFIX}attention_mask" in obs_out
+        assert f"{TOPREWARD_FEATURE_PREFIX}{key}" in obs_out
-    assert f"{TOPREWARD_FEATURE_PREFIX}prompt_length" in obs_out
 
-    prompt_length = obs_out[f"{TOPREWARD_FEATURE_PREFIX}prompt_length"]
+    input_ids = obs_out[f"{TOPREWARD_FEATURE_PREFIX}input_ids"]
-    assert prompt_length.dtype == torch.long
+    labels = obs_out[f"{TOPREWARD_FEATURE_PREFIX}labels"]
-    assert prompt_length.shape == (1,)
+    assert labels.dtype == torch.long
+    assert labels.shape == (2, 10)
+    assert labels[:, :-1].eq(-100).all()
+    assert labels[:, -1].equal(input_ids[:, -1])
 
 
 @_skip_if_topreward_extras_missing
@@ -255,10 +244,3 @@ def test_encoder_step_rejects_missing_image_key(monkeypatch):
     step = _build_step(monkeypatch, image_key="observation.images.top")
     with pytest.raises(KeyError, match="image key"):
         step(_make_transition(observation={}, complementary={"task": "pick"}))
-
-
-@_skip_if_topreward_extras_missing
-def test_encoder_step_rejects_non_dict_observation(monkeypatch):
-    step = _build_step(monkeypatch)
-    with pytest.raises(ValueError, match="observation dict"):
-        step({TransitionKey.OBSERVATION: torch.zeros(1, 3, 8, 8)})

uv.lock

@@ -3010,7 +3010,6 @@ test = [
     { name = "pytest-timeout" },
 ]
 topreward = [
-    { name = "qwen-vl-utils" },
     { name = "transformers" },
 ]
 training = [
@@ -3158,7 +3157,6 @@ requires-dist = [
     { name = "lerobot", extras = ["pyzmq-dep"], marker = "extra == 'unitree-g1'" },
     { name = "lerobot", extras = ["qwen-vl-utils-dep"], marker = "extra == 'eo1'" },
     { name = "lerobot", extras = ["qwen-vl-utils-dep"], marker = "extra == 'sarm'" },
-    { name = "lerobot", extras = ["qwen-vl-utils-dep"], marker = "extra == 'topreward'" },
     { name = "lerobot", extras = ["qwen-vl-utils-dep"], marker = "extra == 'wallx'" },
     { name = "lerobot", extras = ["reachy2"], marker = "extra == 'all'" },
     { name = "lerobot", extras = ["rebot"], marker = "extra == 'all'" },