add training support

lerobot/common/datasets/factory.py

@@ -32,6 +32,7 @@ IMAGENET_STATS = {
     "std": [[[0.229]], [[0.224]], [[0.225]]],  # (c,1,1)
 }
 
+from lerobot.common.datasets.utils_must import (EPISODES_DATASET_MAPPING, TRAINING_FEATURES, FEATURE_KEYS_MAPPING)
 
 def resolve_delta_timestamps(
     cfg: PreTrainedConfig, ds_meta: LeRobotDatasetMetadata
@@ -66,7 +67,7 @@ def resolve_delta_timestamps(
     return delta_timestamps
 
 
-def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDataset:
+def make_dataset1(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDataset:
     """Handles the logic of setting up delta timestamps and image transforms before creating a dataset.
 
     Args:
@@ -116,3 +117,96 @@ def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDatas
                 dataset.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
 
     return dataset
+
+def make_dataset(cfg: TrainPipelineConfig) -> LeRobotDataset | MultiLeRobotDataset:
+    """Handles the logic of setting up delta timestamps and image transforms before creating a dataset.
+
+    Args:
+        cfg (TrainPipelineConfig): A TrainPipelineConfig config which contains a DatasetConfig and a PreTrainedConfig.
+
+    Raises:
+        NotImplementedError: The MultiLeRobotDataset is currently deactivated.
+
+    Returns:
+        LeRobotDataset | MultiLeRobotDataset
+    """
+    image_transforms = (
+        ImageTransforms(cfg.dataset.image_transforms) if cfg.dataset.image_transforms.enable else None
+    )
+    if "," in cfg.dataset.repo_id:
+        repo_id = cfg.dataset.repo_id.split(",")
+        repo_id = [r for r in repo_id if r]
+    else:
+        repo_id = cfg.dataset.repo_id
+    sampling_weights = cfg.dataset.sampling_weights.split(",") if cfg.dataset.sampling_weights else None
+    feature_keys_mapping = FEATURE_KEYS_MAPPING
+    if isinstance(repo_id, str):
+        revision = getattr(cfg.dataset, "revision", None)
+        ds_meta = LeRobotDatasetMetadata(
+            cfg.dataset.repo_id,
+            local_files_only=cfg.dataset.local_files_only,
+            feature_keys_mapping=feature_keys_mapping,
+            revision=revision,
+        )
+        delta_timestamps = resolve_delta_timestamps(cfg.policy, ds_meta)
+        dataset = LeRobotDataset(
+            cfg.dataset.repo_id,
+            root=getattr(cfg.dataset, "root", None),
+            episodes=cfg.dataset.episodes,
+            delta_timestamps=delta_timestamps,
+            image_transforms=image_transforms,
+            revision=revision,
+            video_backend=cfg.dataset.video_backend,
+            local_files_only=cfg.dataset.local_files_only,
+            feature_keys_mapping=feature_keys_mapping,
+            max_action_dim=cfg.dataset.max_action_dim,
+            max_state_dim=cfg.dataset.max_state_dim,
+            max_num_images=cfg.dataset.max_num_images,
+            max_image_dim=cfg.dataset.max_image_dim,
+        )
+    else:
+        delta_timestamps = {}
+        episodes = {}
+        for i in range(len(repo_id)):
+            ds_meta = LeRobotDatasetMetadata(
+                repo_id[i],
+                local_files_only=cfg.dataset.local_files_only,
+                feature_keys_mapping=feature_keys_mapping,
+            )  # FIXME(mshukor): ?
+            delta_timestamps[repo_id[i]] = resolve_delta_timestamps(cfg.policy, ds_meta)
+            episodes[repo_id[i]] =  EPISODES_DATASET_MAPPING.get(repo_id[i], cfg.dataset.episodes)
+        training_features = TRAINING_FEATURES.get(cfg.dataset.features_version, None)
+        dataset = MultiLeRobotDataset(
+            repo_id,
+            # TODO(aliberts): add proper support for multi dataset
+            episodes=episodes,
+            delta_timestamps=delta_timestamps,
+            image_transforms=image_transforms,
+            video_backend=cfg.dataset.video_backend,
+            local_files_only=cfg.dataset.local_files_only,
+            sampling_weights=sampling_weights,
+            feature_keys_mapping=feature_keys_mapping,
+            max_action_dim=cfg.dataset.max_action_dim,
+            max_state_dim=cfg.dataset.max_state_dim,
+            max_num_images=cfg.dataset.max_num_images,
+            max_image_dim=cfg.dataset.max_image_dim,
+            train_on_all_features=cfg.dataset.train_on_all_features,
+            training_features=training_features,
+            discard_first_n_frames=cfg.dataset.discard_first_n_frames,
+            min_fps=cfg.dataset.min_fps,
+            max_fps=cfg.dataset.max_fps,
+            discard_first_idle_frames=cfg.dataset.discard_first_idle_frames,
+            motion_threshold=cfg.dataset.motion_threshold,
+            motion_window_size=cfg.dataset.motion_window_size,
+            motion_buffer=cfg.dataset.motion_buffer,
+        )
+        logging.info(
+            "Multiple datasets were provided. Applied the following index mapping to the provided datasets: "
+            f"{pformat(dataset.repo_id_to_index , indent=2)}"
+        )
+    if cfg.dataset.use_imagenet_stats:
+        for key in dataset.meta.camera_keys:
+            for stats_type, stats in IMAGENET_STATS.items():
+                dataset.meta.stats[key][stats_type] = torch.tensor(stats, dtype=torch.float32)
+
+    return dataset

lerobot/common/datasets/lerobot_dataset.py

@@ -83,6 +83,7 @@ from lerobot.common.datasets.utils_must import (
     create_padded_features,
     pad_tensor,
     map_dict_keys,
+    find_start_of_motion,
     ROBOT_TYPE_KEYS_MAPPING,
     OBS_IMAGE,
     OBS_IMAGE_2,
@@ -562,7 +563,7 @@ class LeRobotDataset(torch.utils.data.Dataset):
                 self.subset_frame_ids += [frame_idx for frame_idx in range(from_ + int(self.fps*self.discard_first_n_frames), to_)]
         elif self.discard_first_idle_frames:
             print(f"Discarding first idle frames: motion_threshold={self.motion_threshold}, motion_window_size={self.motion_window_size}, motion_buffer={self.motion_buffer}")
-            self.robot_states = torch.stack(self.hf_dataset[OBS_ROBOT]).numpy()  # shape: [T, D]
+            self.robot_states = torch.stack(self.hf_dataset[OBS_STATE]).numpy()  # shape: [T, D]
             self.subset_frame_ids = []
             for ep_idx in range(self.num_episodes):
                 from_ = self.episode_data_index["from"][ep_idx]

lerobot/common/datasets/utils_must.py

@@ -191,6 +191,13 @@ def map_dict_keys(item: dict, feature_keys_mapping: dict, training_features: lis
                 features[key] = item[key]
     return features
 
+def find_start_of_motion(velocities, window_size, threshold, motion_buffer):
+    for t in range(len(velocities) - window_size):
+        window_mean = velocities[t:t+window_size].mean()
+        if window_mean > threshold:
+            return max(0, t - motion_buffer)  # include slight context before motion
+    return 0
+
 import yaml
 import requests
 def load_yaml_mapping(name: str) -> dict:
@@ -205,4 +212,19 @@ def load_yaml_mapping(name: str) -> dict:
     return yaml.safe_load(response.text)
 
 # Example usage
-TASKS_KEYS_MAPPING = load_yaml_mapping("features")
+TASKS_KEYS_MAPPING = load_yaml_mapping("tasks")
+FEATURE_KEYS_MAPPING = load_yaml_mapping("features")
+EPISODES_DATASET_MAPPING = {
+    "cadene/droid_1.0.1": list(range(50)),
+    "danaaubakirova/svla_so100_task5_v3": [0, 1, 2, 3, 4, 5, 6, 7, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51],
+    "danaaubakirova/svla_so100_task4_v3": [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17, 18, 19, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31, 32, 33, 34, 35, 40, 41, 42, 43, 44, 45, 46, 47, 48, 49, 50, 51, 52, 53],
+}
+ACTION = "action"
+OBS_STATE = "observation.state"
+TASK = "task"
+ROBOT = "robot_type"
+TRAINING_FEATURES = {
+    0: [ACTION, OBS_STATE, TASK, ROBOT, OBS_IMAGE],
+    1: [ACTION, OBS_STATE, TASK, ROBOT, OBS_IMAGE, OBS_IMAGE_2],
+    2: [ACTION, OBS_STATE, TASK, ROBOT, OBS_IMAGE, OBS_IMAGE_2, OBS_IMAGE_3],
+}

lerobot/common/policies/smolvla2/configuration_smolvla2.py

@@ -0,0 +1,191 @@
+# Copyright 2025 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.
+
+from dataclasses import dataclass, field
+
+from lerobot.common.optim.optimizers import AdamWConfig
+from lerobot.common.optim.schedulers import (
+    CosineDecayWithWarmupSchedulerConfig,
+)
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import FeatureType, NormalizationMode, PolicyFeature
+
+
+@dataclass
+class PEFTConfig:
+    r: int = 4
+    lora_alpha: int = 16
+    lora_dropout: float = 0.1
+    target_modules: str = "q_proj,v_proj"
+
+
+@PreTrainedConfig.register_subclass("smolvla2")
+@dataclass
+class SmolVLA2Config(PreTrainedConfig):
+    # Input / output structure.
+    n_obs_steps: int = 1
+    chunk_size: int = 50
+    n_action_steps: int = 50
+
+    normalization_mapping: dict[str, NormalizationMode] = field(
+        default_factory=lambda: {
+            "VISUAL": NormalizationMode.IDENTITY,
+            "STATE": NormalizationMode.MEAN_STD,
+            "ACTION": NormalizationMode.MEAN_STD,
+        }
+    )
+
+    # Shorter state and action vectors will be padded
+    max_state_dim: int = 32
+    max_action_dim: int = 32
+
+    # Image preprocessing
+    resize_imgs_with_padding: tuple[int, int] = (512, 512)
+
+    # Add empty images. Used by smolvla_aloha_sim which adds the empty
+    # left and right wrist cameras in addition to the top camera.
+    empty_cameras: int = 0
+
+    # Converts the joint and gripper values from the standard Aloha space to
+    # the space used by the pi internal runtime which was used to train the base model.
+    adapt_to_pi_aloha: bool = False
+
+    # Converts joint dimensions to deltas with respect to the current state before passing to the model.
+    # Gripper dimensions will remain in absolute values.
+    use_delta_joint_actions_aloha: bool = False
+
+    # Tokenizer
+    tokenizer_max_length: int = 48
+    proj_width: int = 480
+    # Decoding
+    num_steps: int = 10
+
+    # Attention utils
+    use_cache: bool = True
+
+    # Finetuning settings
+    freeze_vision_encoder: bool = True
+    train_expert_only: bool = False
+    train_state_proj: bool = True
+
+    # Training presets
+    optimizer_lr: float = 2.5e-5  # 1e-4
+    optimizer_betas: tuple[float, float] = (0.9, 0.95)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 1e-10
+    optimizer_grad_clip_norm: float = 10
+    optimizer_lr_vlm: float = 0
+
+    scheduler_warmup_steps: int = 1_000
+    scheduler_decay_steps: int = 30_000
+    scheduler_decay_lr: float = 2.5e-6
+
+    vlm_model_name: str = "HuggingFaceTB/SmolVLM2-500M-Video-Instruct"  # Select the VLM backbone.
+    load_vlm_weights: bool = False  # Set to True in case of training the expert from scratch. True when init from pretrained SmolVLA weights
+    checkpoint_path: str = None
+    peft_method: str = ""
+    peft_config: PEFTConfig = field(default_factory=PEFTConfig)
+    peft_target_model: str = ""
+    add_image_special_tokens: bool = False  # Whether to use special image tokens around image features.
+
+    attention_mode: str = "cross_attn"
+
+    prefix_length: int = -1
+
+    pad_language_to: str = "longest"  # "max_length"
+
+    num_expert_layers: int = -1  # Less or equal to 0 is the default where the action expert has the same number of layers of VLM. Otherwise the expert have less layers.
+    num_vlm_layers: int = 16
+    past_obs_keys: str = "image"
+    add_local_special_image_tokens: bool = False
+
+    reverse_images_order: bool = False
+
+    state_to_prefix: bool = False
+
+    pad_language_to: str = "longest"  # "max_length"
+    causal_action_attention_mask: bool = False
+
+    self_attn_every_n_layers: int = -1  # Number of layers used in the VLM (first num_vlm_layers layers)
+    # self_attn_every_n_layers: int = 2  # Interleave SA layers each self_attn_every_n_layers
+    expert_width_multiplier: float = 0.75  # The action expert hidden size (wrt to the VLM)
+
+    min_period: float = 4e-3  # sensitivity range for the timestep used in sine-cosine positional encoding
+    max_period: float = 4.0
+
+    robot_type: str = ""
+
+    self_attn_only_actions: bool = False
+
+    causal_attention_on_history: bool = False
+
+    predict_relative_actions: bool = False
+    relative_actions_mode: str = "first"
+
+    shuffle_camera_positions: bool = False
+    vlm_img_size: int = -1
+
+    regression_loss: bool = False
+
+    def __post_init__(self):
+        super().__post_init__()
+
+        """Input validation (not exhaustive)."""
+        if self.n_action_steps > self.chunk_size:
+            raise ValueError(
+                f"The chunk size is the upper bound for the number of action steps per model invocation. Got "
+                f"{self.n_action_steps} for `n_action_steps` and {self.chunk_size} for `chunk_size`."
+            )
+        if self.use_delta_joint_actions_aloha:
+            raise NotImplementedError(
+                "`use_delta_joint_actions_aloha` is used by smolvla for aloha real models. It is not ported yet in LeRobot."
+            )
+
+    def validate_features(self) -> None:
+        for i in range(self.empty_cameras):
+            key = f"observation.images.empty_camera_{i}"
+            empty_camera = PolicyFeature(
+                type=FeatureType.VISUAL,
+                shape=(3, 480, 640),
+            )
+            self.input_features[key] = empty_camera
+
+    def get_optimizer_preset(self) -> AdamWConfig:
+        return AdamWConfig(
+            lr=self.optimizer_lr,
+            betas=self.optimizer_betas,
+            eps=self.optimizer_eps,
+            weight_decay=self.optimizer_weight_decay,
+            grad_clip_norm=self.optimizer_grad_clip_norm,
+        )
+
+    def get_scheduler_preset(self):
+        return CosineDecayWithWarmupSchedulerConfig(
+            peak_lr=self.optimizer_lr,
+            decay_lr=self.scheduler_decay_lr,
+            num_warmup_steps=self.scheduler_warmup_steps,
+            num_decay_steps=self.scheduler_decay_steps,
+        )
+
+    @property
+    def observation_delta_indices(self) -> list:
+        return [0]
+
+    @property
+    def action_delta_indices(self) -> list:
+        return list(range(self.chunk_size))
+
+    @property
+    def reward_delta_indices(self) -> None:
+        return None

lerobot/common/policies/smolvla2/smolvlm_with_expert2.py

@@ -0,0 +1,599 @@
+# Copyright 2025 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import copy
+from typing import List, Optional
+
+import torch
+from torch import nn
+from transformers import (
+    AutoConfig,
+    AutoModel,
+    AutoModelForImageTextToText,
+    AutoProcessor,
+    SmolVLMForConditionalGeneration,
+)
+
+
+def apply_rope(x, positions, max_wavelength=10_000):
+    """
+    Applies RoPE positions [B, L] to x [B, L, H, D].
+    """
+    d_half = x.shape[-1] // 2
+    device = x.device
+    dtype = x.dtype
+    x = x.to(torch.float32)
+
+    freq_exponents = (2.0 / x.shape[-1]) * torch.arange(d_half, dtype=torch.float32, device=device)
+    timescale = max_wavelength**freq_exponents
+    radians = positions[..., None].to(torch.float32) / timescale[None, None, :].to(torch.float32)
+
+    radians = radians[..., None, :]
+
+    sin = torch.sin(radians)  # .to(dtype=dtype)
+    cos = torch.cos(radians)  # .to(dtype=dtype)
+
+    x1, x2 = x.split(d_half, dim=-1)
+    res = torch.empty_like(x)
+    res[..., :d_half] = x1 * cos - x2 * sin
+    res[..., d_half:] = x2 * cos + x1 * sin
+
+    return res.to(dtype)
+
+
+def get_intermediate_size(hidden_dim, ffn_dim_multiplier=4, multiple_of=256):
+    hidden_dim = int(2 * hidden_dim / 3)
+    hidden_dim = int(ffn_dim_multiplier * hidden_dim)
+    hidden_dim = multiple_of * ((hidden_dim + multiple_of - 1) // multiple_of)
+    return hidden_dim
+
+
+class SmolVLMWithExpertModel(nn.Module):
+    def __init__(
+        self,
+        model_id: str = "HuggingFaceTB/SmolVLM2-500M-Video-Instruct",
+        load_vlm_weights: bool = True,
+        train_expert_only: bool = True,
+        freeze_vision_encoder: bool = False,
+        attention_mode: str = "self_attn",
+        num_expert_layers: int = -1,
+        num_vlm_layers: int = -1,
+        self_attn_every_n_layers: int = -1,
+        expert_width_multiplier: float = 0.5,
+    ):
+        super().__init__()
+        if load_vlm_weights:
+            print(f"Loading  {model_id} weights ...")
+            self.vlm = AutoModelForImageTextToText.from_pretrained(
+                model_id,
+                device_map="auto",
+                torch_dtype="bfloat16",
+                low_cpu_mem_usage=True,
+            )
+            config = self.vlm.config
+        else:
+            config = AutoConfig.from_pretrained(model_id)
+            self.vlm = SmolVLMForConditionalGeneration(config=config)
+        self.processor = AutoProcessor.from_pretrained(model_id)
+        if num_vlm_layers > 0:
+            print(f"Reducing the number of VLM layers to {num_vlm_layers} ...")
+            self.get_vlm_model().text_model.layers = self.get_vlm_model().text_model.layers[:num_vlm_layers]
+        self.num_vlm_layers = len(self.get_vlm_model().text_model.layers)
+        self.config = config
+        # Smaller lm expert
+        lm_expert_config = copy.deepcopy(config.text_config)
+        hidden_size = lm_expert_config.hidden_size
+        lm_expert_config.hidden_size = int(hidden_size * expert_width_multiplier)  # hidden_size // 2
+        lm_expert_config.intermediate_size = get_intermediate_size(int(hidden_size * expert_width_multiplier))
+        lm_expert_config.num_hidden_layers = self.num_vlm_layers
+        if num_expert_layers > 0:
+            assert len(self.get_vlm_model().text_model.layers) % num_expert_layers == 0, (
+                f"Number of layers in the VLM {len(self.get_vlm_model().text_model.layers)} are not multiple of num_expert_layers {num_expert_layers}"
+            )
+            lm_expert_config.num_hidden_layers = num_expert_layers
+        self.lm_expert = AutoModel.from_config(lm_expert_config)
+
+        self.num_expert_layers = len(self.lm_expert.layers)
+        self.self_attn_every_n_layers = self_attn_every_n_layers
+        if "cross" in attention_mode:
+            # Reshape qkv projections to have the same input dimension as the vlm
+            for layer_idx in range(len(self.lm_expert.layers)):
+                if self.self_attn_every_n_layers > 0 and layer_idx % self.self_attn_every_n_layers == 0:
+                    continue
+                self.lm_expert.layers[layer_idx].self_attn.k_proj = nn.Linear(
+                    config.text_config.num_key_value_heads * config.text_config.head_dim,
+                    lm_expert_config.num_key_value_heads * lm_expert_config.head_dim,
+                    bias=lm_expert_config.attention_bias,
+                )
+                self.lm_expert.layers[layer_idx].self_attn.v_proj = nn.Linear(
+                    config.text_config.num_key_value_heads * config.text_config.head_dim,
+                    lm_expert_config.num_key_value_heads * lm_expert_config.head_dim,
+                    bias=lm_expert_config.attention_bias,
+                )
+        # Remove unused embed_tokens
+        self.lm_expert.embed_tokens = None
+
+        self.num_attention_heads = self.config.text_config.num_attention_heads
+        self.num_key_value_heads = self.config.text_config.num_key_value_heads
+
+        self.freeze_vision_encoder = freeze_vision_encoder
+        self.train_expert_only = train_expert_only
+        self.attention_mode = attention_mode
+        self.expert_hidden_size = lm_expert_config.hidden_size
+        self.set_requires_grad()
+
+    def configure_peft(self, config):
+        # return model
+        self.peft_method = config.peft_method
+        self.peft_target_model = config.peft_target_model
+        if "lora" in self.peft_method:
+            peft_config = config.peft_config
+            target_modules = peft_config.target_modules
+            if not isinstance(target_modules, list):
+                target_modules = target_modules.split(",")
+            lora_config = LoraConfig(
+                task_type=TaskType.CAUSAL_LM,  # Based on the task type (e.g., language modeling, etc.)
+                r=peft_config.r,  # The rank of the low-rank adaptation
+                lora_alpha=peft_config.lora_alpha,  # Scaling factor
+                lora_dropout=peft_config.lora_dropout,  # Dropout applied to LoRA layers
+                target_modules=target_modules,  # The components where LoRA is applied
+                exclude_modules=[
+                    "lm_expert",
+                    "model.lm_expert.model.layers",
+                ],  # FIXME(mshukor): this does not work for now
+            )
+            self.lora_config = lora_config
+            # Apply LoRA and ensure only LoRA parameters are trainable
+            if "text" in self.peft_target_model:
+                self.get_vlm_model().text_model = get_peft_model(self.get_vlm_model().text_model, lora_config)
+            else:
+                self.vlm = get_peft_model(self.vlm, lora_config)
+            # assert config.train_expert_only, "Backbone should be frozen and only lora parameters are " # FIXME(mshukor): handle this here?
+            for name, param in self.vlm.named_parameters():
+                if (
+                    "lora" in name and "text_model.model.layers.17" not in name
+                ):  # lm_head is not a parameter in most LLMs becasue it's tied to the embedding layer
+                    param.requires_grad = True
+                else:
+                    param.requires_grad = False
+
+    def merge_lora_weights(self):
+        """
+        Merge LoRA weights into the base model.
+        """
+        if "text" in self.peft_target_model:
+            self.get_vlm_model().text_model = self.get_vlm_model().text_model.merge_and_unload()
+        else:
+            self.vlm = self.vlm.merge_and_unload()
+
+    def get_vlm_model(
+        self,
+    ):
+        if hasattr(self.vlm.model, "model"):  # When using peft
+            return self.vlm.model.model
+        else:
+            return self.vlm.model
+
+    def set_requires_grad(self):
+        if self.freeze_vision_encoder:
+            self.get_vlm_model().vision_model.eval()
+            for params in self.get_vlm_model().vision_model.parameters():
+                params.requires_grad = False
+        if self.train_expert_only:
+            self.vlm.eval()
+            for params in self.vlm.parameters():
+                params.requires_grad = False
+        else:
+            # To avoid unused params issue with distributed training
+            last_layers = [self.num_vlm_layers - 1]
+            if (
+                self.num_vlm_layers != self.num_expert_layers
+                and self.num_vlm_layers % self.num_expert_layers == 0
+            ):
+                last_layers.append(self.num_vlm_layers - 2)
+            frozen_layers = [
+                "lm_head",
+                "text_model.model.norm.weight",
+            ]
+            for layer in last_layers:
+                frozen_layers.append(f"text_model.model.layers.{layer}.")
+
+            for name, params in self.vlm.named_parameters():
+                if any(k in name for k in frozen_layers):
+                    params.requires_grad = False
+        # To avoid unused params issue with distributed training
+        for name, params in self.lm_expert.named_parameters():
+            if "lm_head" in name:
+                params.requires_grad = False
+
+    def train(self, mode: bool = True):
+        super().train(mode)
+
+        if self.freeze_vision_encoder:
+            self.get_vlm_model().vision_model.eval()
+
+        if self.train_expert_only:
+            self.vlm.eval()
+
+    def embed_image(self, image: torch.Tensor):
+        patch_attention_mask = None
+        # Get sequence from the vision encoder
+        image_hidden_states = (
+            self.get_vlm_model()
+            .vision_model(
+                pixel_values=image.to(dtype=self.get_vlm_model().vision_model.dtype),
+                patch_attention_mask=patch_attention_mask,
+            )
+            .last_hidden_state
+        )
+        # Modality projection & resampling
+        image_hidden_states = self.get_vlm_model().connector(image_hidden_states)
+        return image_hidden_states
+
+    def embed_language_tokens(self, tokens: torch.Tensor):
+        return self.get_vlm_model().text_model.get_input_embeddings()(tokens)
+
+    def forward_attn_layer(
+        self,
+        model_layers,
+        inputs_embeds,
+        layer_idx,
+        position_ids,
+        attention_mask,
+        batch_size,
+        head_dim,
+        use_cache: bool = True,
+        fill_kv_cache: bool = True,
+        past_key_values=None,
+    ) -> list[torch.Tensor]:
+        query_states = []
+        key_states = []
+        value_states = []
+        for i, hidden_states in enumerate(inputs_embeds):
+            layer = model_layers[i][layer_idx]
+            if hidden_states is None or layer is None:
+                continue
+            hidden_states = layer.input_layernorm(hidden_states)
+
+            input_shape = hidden_states.shape[:-1]
+            hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
+
+            hidden_states = hidden_states.to(dtype=layer.self_attn.q_proj.weight.dtype)
+            query_state = layer.self_attn.q_proj(hidden_states).view(hidden_shape)
+            key_state = layer.self_attn.k_proj(hidden_states).view(hidden_shape)
+            value_state = layer.self_attn.v_proj(hidden_states).view(hidden_shape)
+
+            query_states.append(query_state)
+            key_states.append(key_state)
+            value_states.append(value_state)
+
+        # B,L,H,D with L sequence length, H number of heads, D head dim
+        # concatenate on the number of embeddings/tokens
+        query_states = torch.cat(query_states, dim=1)
+        key_states = torch.cat(key_states, dim=1)
+        value_states = torch.cat(value_states, dim=1)
+        seq_len = query_states.shape[1]
+        if seq_len < position_ids.shape[1]:
+            _position_ids = position_ids[:, :seq_len]
+            _attention_mask = attention_mask[:, :seq_len, :seq_len]
+        else:
+            _position_ids = position_ids
+            _attention_mask = attention_mask
+
+        attention_mask_ = _attention_mask
+        position_ids_ = _position_ids
+
+        query_states = apply_rope(query_states, position_ids_)
+        key_states = apply_rope(key_states, position_ids_)
+
+        if use_cache and past_key_values is None:
+            past_key_values = {}
+
+        if use_cache:
+            if fill_kv_cache:
+                past_key_values[layer_idx] = {
+                    "key_states": key_states,
+                    "value_states": value_states,
+                }
+            else:
+                # TODO here, some optimization can be done - similar to a `StaticCache` we can declare the `max_len` before.
+                # so we create an empty cache, with just one cuda malloc, and if (in autoregressive case) we reach
+                # the max len, then we (for instance) double the cache size. This implementation already exists
+                # in `transformers`. (molbap)
+                key_states = torch.cat([past_key_values[layer_idx]["key_states"], key_states], dim=1)
+                value_states = torch.cat([past_key_values[layer_idx]["value_states"], value_states], dim=1)
+
+        attention_interface = self.get_attention_interface()
+
+        att_output = attention_interface(
+            attention_mask_, batch_size, head_dim, query_states, key_states, value_states
+        )
+        return [att_output], past_key_values
+
+    def forward_cross_attn_layer(
+        self,
+        model_layers,
+        inputs_embeds,
+        layer_idx,
+        position_ids,
+        attention_mask,
+        batch_size,
+        head_dim,
+        use_cache: bool = True,
+        fill_kv_cache: bool = True,
+        past_key_values=None,
+    ) -> list[torch.Tensor]:
+        attention_interface = self.get_attention_interface()
+
+        att_outputs = []
+        assert len(inputs_embeds) == 2 or (use_cache and past_key_values is not None and not fill_kv_cache), (
+            f"Both len(inputs_embeds) == {len(inputs_embeds)} and past_key_values is {past_key_values}"
+        )
+
+        if len(inputs_embeds) == 2 and not past_key_values:
+            # Prefix attention
+            seq_len = inputs_embeds[0].shape[1]
+            position_id, expert_position_id = position_ids[:, :seq_len], position_ids[:, seq_len:]
+            prefix_attention_mask = attention_mask[:, :seq_len, :seq_len]
+
+            layer = model_layers[0][layer_idx]
+
+            hidden_states = layer.input_layernorm(inputs_embeds[0])
+
+            input_shape = hidden_states.shape[:-1]
+            hidden_shape = (*input_shape, -1, layer.self_attn.head_dim)
+
+            hidden_states = hidden_states.to(dtype=layer.self_attn.q_proj.weight.dtype)
+            query_state = layer.self_attn.q_proj(hidden_states).view(hidden_shape)
+            key_state = layer.self_attn.k_proj(hidden_states).view(hidden_shape)
+            value_states = layer.self_attn.v_proj(hidden_states).view(hidden_shape)
+
+            # B,L,H,D with L sequence length, H number of heads, D head dim
+            query_states = apply_rope(query_state, position_id)
+            key_states = apply_rope(key_state, position_id)
+
+            att_output = attention_interface(
+                prefix_attention_mask, batch_size, head_dim, query_states, key_states, value_states
+            )
+            att_outputs.append(att_output)
+        else:
+            expert_position_id = position_ids
+
+        if use_cache and past_key_values is None:
+            past_key_values = {}
+
+        if use_cache:
+            if fill_kv_cache:
+                past_key_values[layer_idx] = {
+                    "key_states": key_states,
+                    "value_states": value_states,
+                }
+            else:
+                # TODO here, some optimization can be done - similar to a `StaticCache` we can declare the `max_len` before.
+                # so we create an empty cache, with just one cuda malloc, and if (in autoregressive case) we reach
+                # the max len, then we (for instance) double the cache size. This implementation already exists
+                # in `transformers`. (molbap)
+                key_states = past_key_values[layer_idx]["key_states"]
+                value_states = past_key_values[layer_idx]["value_states"]
+
+        # Expert
+        expert_layer = model_layers[1][layer_idx]
+        if expert_layer is not None:
+            expert_hidden_states = expert_layer.input_layernorm(inputs_embeds[1])
+
+            expert_input_shape = expert_hidden_states.shape[:-1]
+            expert_hidden_shape = (*expert_input_shape, -1, expert_layer.self_attn.head_dim)
+
+            expert_hidden_states = expert_hidden_states.to(dtype=expert_layer.self_attn.q_proj.weight.dtype)
+            expert_query_state = expert_layer.self_attn.q_proj(expert_hidden_states).view(expert_hidden_shape)
+
+            _key_states = key_states.to(dtype=expert_layer.self_attn.k_proj.weight.dtype).view(
+                *key_states.shape[:2], -1
+            )
+            expert_key_states = expert_layer.self_attn.k_proj(_key_states).view(
+                *_key_states.shape[:-1], -1, expert_layer.self_attn.head_dim
+            )  # k_proj should have same dim as kv
+
+            _value_states = value_states.to(dtype=expert_layer.self_attn.v_proj.weight.dtype).view(
+                *value_states.shape[:2], -1
+            )
+            expert_value_states = expert_layer.self_attn.v_proj(_value_states).view(
+                *_value_states.shape[:-1], -1, expert_layer.self_attn.head_dim
+            )
+
+            expert_position_id = (
+                expert_position_id - torch.min(expert_position_id, dim=1, keepdim=True).values
+            )  # start from 0
+            expert_attention_mask = attention_mask[
+                :, -inputs_embeds[1].shape[1] :, : expert_key_states.shape[1] :
+            ]  # take into account kv
+
+            expert_query_states = apply_rope(expert_query_state, expert_position_id)
+
+            att_output = attention_interface(
+                expert_attention_mask,
+                batch_size,
+                head_dim,
+                expert_query_states,
+                expert_key_states,
+                expert_value_states,
+            )
+            att_outputs.append(att_output)
+        else:
+            att_outputs.append(None)
+
+        # att_output = att_output.to(dtype=models[i].dtype)
+        return att_outputs, past_key_values
+
+    def get_model_layers(self, models: list) -> list:
+        vlm_layers = []
+        expert_layers = []
+        multiple_of = self.num_vlm_layers // self.num_expert_layers
+        for i in range(self.num_vlm_layers):
+            if multiple_of > 0 and i > 0 and i % multiple_of != 0:
+                expert_layer = None
+            else:
+                expert_layer_index = i // multiple_of if multiple_of > 0 else i
+                expert_layer = models[1].layers[expert_layer_index]
+            vlm_layers.append(models[0].layers[i])
+            expert_layers.append(expert_layer)
+        return [vlm_layers, expert_layers]
+
+    def forward(
+        self,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: List[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        fill_kv_cache: Optional[bool] = None,
+    ):
+        models = [self.get_vlm_model().text_model, self.lm_expert]
+        model_layers = self.get_model_layers(models)
+        for hidden_states in inputs_embeds:
+            # TODO this is very inefficient
+            # dtype is always the same, batch size too (if > 1 len)
+            # device could be trickier in multi gpu edge cases but that's it
+            if hidden_states is None:
+                continue
+            batch_size = hidden_states.shape[0]
+
+        # RMSNorm
+        num_layers = self.num_vlm_layers
+        head_dim = self.vlm.config.text_config.head_dim
+        for layer_idx in range(num_layers):
+            if (
+                fill_kv_cache
+                or "cross" not in self.attention_mode
+                or (self.self_attn_every_n_layers > 0 and layer_idx % self.self_attn_every_n_layers == 0)
+            ):
+                att_outputs, past_key_values = self.forward_attn_layer(
+                    model_layers,
+                    inputs_embeds,
+                    layer_idx,
+                    position_ids,
+                    attention_mask,
+                    batch_size,
+                    head_dim,
+                    use_cache=use_cache,
+                    fill_kv_cache=fill_kv_cache,
+                    past_key_values=past_key_values,
+                )
+            else:
+                att_outputs, past_key_values = self.forward_cross_attn_layer(
+                    model_layers,
+                    inputs_embeds,
+                    layer_idx,
+                    position_ids,
+                    attention_mask,
+                    batch_size,
+                    head_dim,
+                    use_cache=use_cache,
+                    fill_kv_cache=fill_kv_cache,
+                    past_key_values=past_key_values,
+                )
+            outputs_embeds = []
+            start = 0
+            for i, hidden_states in enumerate(inputs_embeds):
+                layer = model_layers[i][layer_idx]
+                att_output = (
+                    att_outputs[i] if i < len(att_outputs) else att_outputs[0]
+                )  # in case of self_attn
+                if hidden_states is not None:
+                    if layer is None:
+                        outputs_embeds.append(hidden_states)
+                        continue
+                    end = start + hidden_states.shape[1]
+
+                    if att_output.dtype != layer.self_attn.o_proj.weight.dtype:
+                        att_output = att_output.to(layer.self_attn.o_proj.weight.dtype)
+                    att_out = att_output[:, start:end]
+                    out_emb = layer.self_attn.o_proj(att_out)
+
+                    out_emb += hidden_states
+                    after_first_residual = out_emb.clone()
+
+                    out_emb = layer.post_attention_layernorm(out_emb)
+                    out_emb = layer.mlp(out_emb)
+
+                    out_emb += after_first_residual
+
+                    outputs_embeds.append(out_emb)
+
+                    start = end if len(att_outputs) == 1 else 0
+                else:
+                    outputs_embeds.append(None)
+
+            inputs_embeds = outputs_embeds
+
+        # final norm
+        outputs_embeds = []
+        for i, hidden_states in enumerate(inputs_embeds):
+            if hidden_states is not None:
+                out_emb = models[i].norm(hidden_states)
+                outputs_embeds.append(out_emb)
+            else:
+                outputs_embeds.append(None)
+        return outputs_embeds, past_key_values
+
+    def get_attention_interface(self):
+        attention_interface = self.eager_attention_forward
+        return attention_interface
+
+    def eager_attention_forward(
+        self, attention_mask, batch_size, head_dim, query_states, key_states, value_states
+    ):
+        num_att_heads = self.num_attention_heads
+        num_key_value_heads = self.num_key_value_heads
+        num_key_value_groups = num_att_heads // num_key_value_heads
+
+        sequence_length = key_states.shape[1]
+
+        key_states = key_states[:, :, :, None, :].expand(
+            batch_size, sequence_length, num_key_value_heads, num_key_value_groups, head_dim
+        )
+        key_states = key_states.reshape(
+            batch_size, sequence_length, num_key_value_heads * num_key_value_groups, head_dim
+        )
+
+        value_states = value_states[:, :, :, None, :].expand(
+            batch_size, sequence_length, num_key_value_heads, num_key_value_groups, head_dim
+        )
+        value_states = value_states.reshape(
+            batch_size, sequence_length, num_key_value_heads * num_key_value_groups, head_dim
+        )
+
+        # Attention here is upcasted to float32 to match the original eager implementation.
+        query_states = query_states.to(dtype=torch.float32)
+        key_states = key_states.to(dtype=torch.float32)
+
+        query_states = query_states.transpose(1, 2)
+        key_states = key_states.transpose(1, 2)
+
+        att_weights = torch.matmul(query_states, key_states.transpose(2, 3))
+        att_weights *= head_dim**-0.5
+
+        att_weights = att_weights.to(dtype=torch.float32)
+        big_neg = torch.finfo(att_weights.dtype).min  # -2.3819763e38  # See gemma/modules.py
+        masked_att_weights = torch.where(attention_mask[:, None, :, :], att_weights, big_neg)
+        probs = nn.functional.softmax(masked_att_weights, dim=-1)
+        probs = probs.to(dtype=value_states.dtype)
+
+        att_output = torch.matmul(probs, value_states.permute(0, 2, 1, 3))
+
+        att_output = att_output.permute(0, 2, 1, 3)
+        # we use -1 because sequence length can change
+        att_output = att_output.reshape(batch_size, -1, num_key_value_heads * num_key_value_groups * head_dim)
+
+        return att_output