Add Pi0 (#681)

Co-authored-by: Simon Alibert <simon.alibert@huggingface.co>
Co-authored-by: Simon Alibert <75076266+aliberts@users.noreply.github.com>
Co-authored-by: Pablo <pablo.montalvo.leroux@gmail.com>

lerobot/common/datasets/lerobot_dataset.py

@@ -672,6 +672,10 @@ class LeRobotDataset(torch.utils.data.Dataset):
             for cam in image_keys:
                 item[cam] = self.image_transforms(item[cam])
 
+        # Add task as a string
+        task_idx = item["task_index"].item()
+        item["task"] = self.meta.tasks[task_idx]
+
         return item
 
     def __repr__(self):

lerobot/common/optim/optimizers.py

@@ -8,8 +8,6 @@ import torch
 @dataclass
 class OptimizerConfig(draccus.ChoiceRegistry, abc.ABC):
     lr: float
-    betas: tuple[float, float]
-    eps: float
     weight_decay: float
     grad_clip_norm: float
 
@@ -54,3 +52,19 @@ class AdamWConfig(OptimizerConfig):
         kwargs = asdict(self)
         kwargs.pop("grad_clip_norm")
         return torch.optim.AdamW(params, **kwargs)
+
+
+@OptimizerConfig.register_subclass("sgd")
+@dataclass
+class SGDConfig(OptimizerConfig):
+    lr: float = 1e-3
+    momentum: float = 0.0
+    dampening: float = 0.0
+    nesterov: bool = False
+    weight_decay: float = 0.0
+    grad_clip_norm: float = 10.0
+
+    def build(self, params: dict) -> torch.optim.Optimizer:
+        kwargs = asdict(self)
+        kwargs.pop("grad_clip_norm")
+        return torch.optim.SGD(params, **kwargs)

lerobot/common/optim/schedulers.py

@@ -54,3 +54,38 @@ class VQBeTSchedulerConfig(LRSchedulerConfig):
                 return max(0.0, 0.5 * (1.0 + math.cos(math.pi * float(self.num_cycles) * 2.0 * progress)))
 
         return LambdaLR(optimizer, lr_lambda, -1)
+
+
+@LRSchedulerConfig.register_subclass("cosine_decay_with_warmup")
+@dataclass
+class CosineDecayWithWarmupSchedulerConfig(LRSchedulerConfig):
+    """Used by Physical Intelligence to train Pi0"""
+
+    num_warmup_steps: int
+    num_decay_steps: int
+    peak_lr: float
+    decay_lr: float
+
+    def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
+        del num_training_steps
+
+        def lr_lambda(current_step):
+            def linear_warmup_schedule(current_step):
+                if current_step <= 0:
+                    return 1 / (self.num_warmup_steps + 1)
+                frac = 1 - current_step / self.num_warmup_steps
+                return (1 / (self.num_warmup_steps + 1) - 1) * frac + 1
+
+            def cosine_decay_schedule(current_step):
+                step = min(current_step, self.num_decay_steps)
+                cosine_decay = 0.5 * (1 + math.cos(math.pi * step / self.num_decay_steps))
+                alpha = self.decay_lr / self.peak_lr
+                decayed = (1 - alpha) * cosine_decay + alpha
+                return decayed
+
+            if current_step < self.num_warmup_steps:
+                return linear_warmup_schedule(current_step)
+
+            return cosine_decay_schedule(current_step)
+
+        return LambdaLR(optimizer, lr_lambda, -1)

lerobot/common/policies/__init__.py

@@ -1,4 +1,5 @@
 from .act.configuration_act import ACTConfig as ACTConfig
 from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
+from .pi0.configuration_pi0 import PI0Config as PI0Config
 from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
 from .vqbet.configuration_vqbet import VQBeTConfig as VQBeTConfig

lerobot/common/policies/factory.py

@@ -25,6 +25,7 @@ from lerobot.common.envs.configs import EnvConfig
 from lerobot.common.envs.utils import env_to_policy_features
 from lerobot.common.policies.act.configuration_act import ACTConfig
 from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
+from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
 from lerobot.common.policies.pretrained import PreTrainedPolicy
 from lerobot.common.policies.tdmpc.configuration_tdmpc import TDMPCConfig
 from lerobot.common.policies.vqbet.configuration_vqbet import VQBeTConfig
@@ -50,6 +51,10 @@ def get_policy_class(name: str) -> PreTrainedPolicy:
         from lerobot.common.policies.vqbet.modeling_vqbet import VQBeTPolicy
 
         return VQBeTPolicy
+    elif name == "pi0":
+        from lerobot.common.policies.pi0.modeling_pi0 import PI0Policy
+
+        return PI0Policy
     else:
         raise NotImplementedError(f"Policy with name {name} is not implemented.")
 
@@ -63,6 +68,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
         return ACTConfig(**kwargs)
     elif policy_type == "vqbet":
         return VQBeTConfig(**kwargs)
+    elif policy_type == "pi0":
+        return PI0Config(**kwargs)
     else:
         raise ValueError(f"Policy type '{policy_type}' is not available.")
 
@@ -141,4 +148,6 @@ def make_policy(
     policy.to(device)
     assert isinstance(policy, nn.Module)
 
+    # policy = torch.compile(policy, mode="reduce-overhead")
+
     return policy

lerobot/common/policies/normalize.py

@@ -140,6 +140,9 @@ class Normalize(nn.Module):
     def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
         batch = dict(batch)  # shallow copy avoids mutating the input batch
         for key, ft in self.features.items():
+            if key not in batch:
+                continue
+
             norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
             if norm_mode is NormalizationMode.IDENTITY:
                 continue
@@ -210,6 +213,9 @@ class Unnormalize(nn.Module):
     def forward(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
         batch = dict(batch)  # shallow copy avoids mutating the input batch
         for key, ft in self.features.items():
+            if key not in batch:
+                continue
+
             norm_mode = self.norm_map.get(ft.type, NormalizationMode.IDENTITY)
             if norm_mode is NormalizationMode.IDENTITY:
                 continue

lerobot/common/policies/pi0/configuration_pi0.py

@@ -0,0 +1,134 @@
+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
+
+
+@PreTrainedConfig.register_subclass("pi0")
+@dataclass
+class PI0Config(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] = (224, 224)
+
+    # Add empty images. Used by pi0_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
+
+    # Projector
+    proj_width: int = 1024
+
+    # Decoding
+    num_steps: int = 10
+
+    # Attention utils
+    use_cache: bool = True
+    attention_implementation: str = "eager"  # or fa2, flex
+
+    # 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
+    optimizer_betas: tuple[float, float] = (0.9, 0.95)
+    optimizer_eps: float = 1e-8
+    optimizer_weight_decay: float = 1e-10
+
+    scheduler_warmup_steps: int = 1_000
+    scheduler_decay_steps: int = 30_000
+    scheduler_decay_lr: float = 2.5e-6
+
+    # TODO: Add EMA
+
+    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.n_obs_steps != 1:
+            raise ValueError(
+                f"Multiple observation steps not handled yet. Got `nobs_steps={self.n_obs_steps}`"
+            )
+
+        if self.use_delta_joint_actions_aloha:
+            raise NotImplementedError(
+                "`use_delta_joint_actions_aloha` is used by pi0 for aloha real models. It is not ported yet in LeRobot."
+            )
+
+    def validate_features(self) -> None:
+        # TODO: implement value error
+        # if not self.image_features and not self.env_state_feature:
+        #     raise ValueError("You must provide at least one image or the environment state among the inputs.")
+
+        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,
+        )
+
+    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) -> None:
+        return None
+
+    @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/pi0/conversion_scripts/benchmark.py

@@ -0,0 +1,68 @@
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDataset
+from lerobot.common.policies.factory import make_policy
+from lerobot.configs.policies import PreTrainedConfig
+
+torch.backends.cudnn.benchmark = True
+
+
+def main():
+    device = "cuda"
+    dataset_repo_id = "danaaubakirova/koch_test"
+    # model_name = "pi0_base"
+    # ckpt_torch_dir = Path.home() / f".cache/openpi/openpi-assets/checkpoints/{model_name}_pytorch"
+    ckpt_torch_dir = "lerobot/pi0"
+
+    dataset = LeRobotDataset(dataset_repo_id, episodes=[0])
+
+    dataloader = torch.utils.data.DataLoader(
+        dataset,
+        num_workers=0,
+        batch_size=1,
+    )
+
+    batch = next(iter(dataloader))
+
+    # To device
+    for k in batch:
+        if isinstance(batch[k], torch.Tensor):
+            batch[k] = batch[k].to(device=device, dtype=torch.float32)
+
+    cfg = PreTrainedConfig.from_pretrained(ckpt_torch_dir)
+    cfg.pretrained_path = ckpt_torch_dir
+    policy = make_policy(cfg, device, ds_meta=dataset.meta)
+
+    # policy = torch.compile(policy, mode="reduce-overhead")
+
+    warmup_iters = 10
+    benchmark_iters = 30
+
+    # Warmup
+    for _ in range(warmup_iters):
+        torch.cuda.synchronize()
+        policy.select_action(batch)
+        policy.reset()
+        torch.cuda.synchronize()
+
+    # Benchmark
+    start_event = torch.cuda.Event(enable_timing=True)
+    end_event = torch.cuda.Event(enable_timing=True)
+
+    start_event.record()
+    for _ in range(benchmark_iters):
+        policy.select_action(batch)
+        policy.reset()
+    end_event.record()
+
+    # Synchronize and measure time
+    torch.cuda.synchronize()
+    elapsed_time_ms = start_event.elapsed_time(end_event)
+
+    avg_time_per_iter = elapsed_time_ms / benchmark_iters
+    print(f"Average execution time per iteration: {avg_time_per_iter:.3f} ms")
+
+
+if __name__ == "__main__":
+    with torch.inference_mode():
+        main()

lerobot/common/policies/pi0/conversion_scripts/compare_with_jax.py

@@ -0,0 +1,117 @@
+import json
+import pickle
+from pathlib import Path
+
+import torch
+
+from lerobot.common.datasets.lerobot_dataset import LeRobotDatasetMetadata
+from lerobot.common.policies.factory import make_policy
+from lerobot.configs.policies import PreTrainedConfig
+
+
+def display(tensor: torch.Tensor):
+    if tensor.dtype == torch.bool:
+        tensor = tensor.float()
+    print(f"Shape: {tensor.shape}")
+    print(f"Mean: {tensor.mean().item()}")
+    print(f"Std: {tensor.std().item()}")
+    print(f"Min: {tensor.min().item()}")
+    print(f"Max: {tensor.max().item()}")
+
+
+def main():
+    num_motors = 14
+    device = "cuda"
+    # model_name = "pi0_aloha_towel"
+    model_name = "pi0_aloha_sim"
+
+    if model_name == "pi0_aloha_towel":
+        dataset_repo_id = "lerobot/aloha_static_towel"
+    else:
+        dataset_repo_id = "lerobot/aloha_sim_transfer_cube_human"
+
+    ckpt_torch_dir = Path.home() / f".cache/openpi/openpi-assets/checkpoints/{model_name}_pytorch"
+    ckpt_jax_dir = Path.home() / f".cache/openpi/openpi-assets/checkpoints/{model_name}"
+    save_dir = Path(f"../openpi/data/{model_name}/save")
+
+    with open(save_dir / "example.pkl", "rb") as f:
+        example = pickle.load(f)
+    with open(save_dir / "outputs.pkl", "rb") as f:
+        outputs = pickle.load(f)
+    with open(save_dir / "noise.pkl", "rb") as f:
+        noise = pickle.load(f)
+
+    with open(ckpt_jax_dir / "assets/norm_stats.json") as f:
+        norm_stats = json.load(f)
+
+    # Override stats
+    dataset_meta = LeRobotDatasetMetadata(dataset_repo_id)
+    dataset_meta.stats["observation.state"]["mean"] = torch.tensor(
+        norm_stats["norm_stats"]["state"]["mean"][:num_motors], dtype=torch.float32
+    )
+    dataset_meta.stats["observation.state"]["std"] = torch.tensor(
+        norm_stats["norm_stats"]["state"]["std"][:num_motors], dtype=torch.float32
+    )
+
+    # Create LeRobot batch from Jax
+    batch = {}
+    for cam_key, uint_chw_array in example["images"].items():
+        batch[f"observation.images.{cam_key}"] = torch.from_numpy(uint_chw_array) / 255.0
+    batch["observation.state"] = torch.from_numpy(example["state"])
+    batch["action"] = torch.from_numpy(outputs["actions"])
+    batch["task"] = example["prompt"]
+
+    if model_name == "pi0_aloha_towel":
+        del batch["observation.images.cam_low"]
+    elif model_name == "pi0_aloha_sim":
+        batch["observation.images.top"] = batch["observation.images.cam_high"]
+        del batch["observation.images.cam_high"]
+
+    # Batchify
+    for key in batch:
+        if isinstance(batch[key], torch.Tensor):
+            batch[key] = batch[key].unsqueeze(0)
+        elif isinstance(batch[key], str):
+            batch[key] = [batch[key]]
+        else:
+            raise ValueError(f"{key}, {batch[key]}")
+
+    # To device
+    for k in batch:
+        if isinstance(batch[k], torch.Tensor):
+            batch[k] = batch[k].to(device=device, dtype=torch.float32)
+
+    noise = torch.from_numpy(noise).to(device=device, dtype=torch.float32)
+
+    from lerobot.common import policies  # noqa
+
+    cfg = PreTrainedConfig.from_pretrained(ckpt_torch_dir)
+    cfg.pretrained_path = ckpt_torch_dir
+    policy = make_policy(cfg, device, dataset_meta)
+
+    # loss_dict = policy.forward(batch, noise=noise, time=time_beta)
+    # loss_dict["loss"].backward()
+    # print("losses")
+    # display(loss_dict["losses_after_forward"])
+    # print("pi_losses")
+    # display(pi_losses)
+
+    actions = []
+    for _ in range(50):
+        action = policy.select_action(batch, noise=noise)
+        actions.append(action)
+
+    actions = torch.stack(actions, dim=1)
+    pi_actions = batch["action"]
+    print("actions")
+    display(actions)
+    print()
+    print("pi_actions")
+    display(pi_actions)
+    print("atol=3e-2", torch.allclose(actions, pi_actions, atol=3e-2))
+    print("atol=2e-2", torch.allclose(actions, pi_actions, atol=2e-2))
+    print("atol=1e-2", torch.allclose(actions, pi_actions, atol=1e-2))
+
+
+if __name__ == "__main__":
+    main()

lerobot/common/policies/pi0/conversion_scripts/conversion_utils.py

@@ -0,0 +1,70 @@
+from transformers import GemmaConfig, PaliGemmaConfig
+
+
+def get_paligemma_config(precision: str):
+    config = {
+        "image_token_index": None,
+        "pad_token_id": 0,
+        "bos_token_id": 2,
+        "eos_token_id": 1,
+    }
+
+    # image_sizes = {"2b-test": 224, "3b-224px": 224, "3b-448px": 448, "3b-896px": 896}
+
+    image_size = 224  # image_sizes[variant]
+    patch_size = 14
+    num_image_tokens = (image_size**2) // (patch_size**2)
+
+    config["image_token_index"] = 257152
+    text_config = {
+        "vocab_size": 257152,
+        "num_hidden_layers": 18,
+        "num_key_value_heads": 1,
+        "head_dim": 256,
+        "torch_dtype": precision,
+        "hidden_size": 2048,
+        "hidden_activation": "gelu_pytorch_tanh",
+        "num_attention_heads": 8,
+        "intermediate_size": 16384,
+        "is_encoder_decoder": False,
+    }
+    vision_config = {
+        "torch_dtype": precision,
+        "image_size": image_size,
+        "patch_size": patch_size,
+        "num_image_tokens": num_image_tokens,
+        "hidden_size": 1152,
+        "intermediate_size": 4304,
+        "num_hidden_layers": 27,
+        "num_attention_heads": 16,
+        "projector_hidden_act": "gelu_fast",
+        "vision_use_head": False,
+    }
+    final_config = PaliGemmaConfig(text_config=text_config, vision_config=vision_config, **config)
+    return final_config
+
+
+def get_gemma_config(precision: str):
+    config = {
+        "image_token_index": None,
+        "pad_token_id": 0,
+        "bos_token_id": 2,
+        "eos_token_id": 1,
+    }
+
+    config["image_token_index"] = 257152
+    text_config = {
+        "vocab_size": 257152,
+        "num_hidden_layers": 18,
+        "num_key_value_heads": 1,
+        "head_dim": 256,
+        "torch_dtype": precision,
+        "hidden_size": 1024,
+        "hidden_activation": "gelu_pytorch_tanh",
+        "num_attention_heads": 8,
+        "intermediate_size": 4096,
+        "is_encoder_decoder": False,
+    }
+    final_config = GemmaConfig()
+    final_config.update(text_config)
+    return final_config

lerobot/common/policies/pi0/conversion_scripts/convert_pi0_to_hf_lerobot.py

@@ -0,0 +1,423 @@
+"""
+Convert pi0 parameters from Jax to Pytorch
+
+Follow [README of openpi](https://github.com/Physical-Intelligence/openpi) to create a new environment
+and install the required librairies.
+
+```bash
+cd ~/code/openpi
+source .venv/bin/activate
+```
+
+Example downloading parameters:
+```bash
+python
+>>> import openpi.shared.download as download
+>>> path='s3://openpi-assets/checkpoints/pi0_base/params'
+>>> download.maybe_download(path)
+```
+
+Converting pi0_base:
+```python
+python lerobot/common/policies/pi0/conversion_scripts/convert_pi0_to_hf_lerobot.py \
+    --checkpoint_dir /home/remi_cadene/.cache/openpi/openpi-assets/checkpoints/pi0_base/params \
+    --output_path /home/remi_cadene/.cache/openpi/openpi-assets/checkpoints/pi0_base_pytorch
+```
+
+```python
+python lerobot/common/policies/pi0/conversion_scripts/convert_pi0_to_hf_lerobot.py \
+    --checkpoint_dir /home/remi_cadene/.cache/openpi/openpi-assets/checkpoints/pi0_aloha_sim/params \
+    --output_path /home/remi_cadene/.cache/openpi/openpi-assets/checkpoints/pi0_aloha_sim_pytorch
+```
+"""
+
+import argparse
+import pathlib
+
+import jax
+import numpy as np
+import orbax.checkpoint as ocp
+import torch
+from jax.sharding import SingleDeviceSharding
+
+from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
+from lerobot.common.policies.pi0.conversion_scripts.conversion_utils import (
+    get_gemma_config,
+    get_paligemma_config,
+)
+from lerobot.common.policies.pi0.modeling_pi0 import PI0Policy
+
+PRECISIONS = {"bfloat16": torch.bfloat16, "float32": torch.float32, "float16": torch.float16}
+
+
+def slice_paligemma_state_dict(state_dict, config):
+    suffix = "/value" if "img/embedding/kernel/value" in state_dict else ""
+
+    # fmt: off
+    # patch embeddings
+    state_dict["paligemma.vision_tower.vision_model.embeddings.patch_embedding.weight"] = state_dict.pop(f"img/embedding/kernel{suffix}").transpose(
+        3, 2, 0, 1
+    )
+    state_dict["paligemma.vision_tower.vision_model.embeddings.patch_embedding.bias"] = state_dict.pop(f"img/embedding/bias{suffix}")
+    # positional embeddings
+    state_dict["paligemma.vision_tower.vision_model.embeddings.position_embedding.weight"] = state_dict.pop(f"img/pos_embedding{suffix}").reshape(
+        -1, config.vision_config.hidden_size
+    )
+
+    # extract vision layers to be sliced at index 0. There are 27 layers in the base model.
+    encoderblock_layernorm0_scale = state_dict.pop(f"img/Transformer/encoderblock/LayerNorm_0/scale{suffix}")
+    encoderblock_layernorm0_bias = state_dict.pop(f"img/Transformer/encoderblock/LayerNorm_0/bias{suffix}")
+    encoderblock_layernorm1_scale = state_dict.pop(f"img/Transformer/encoderblock/LayerNorm_1/scale{suffix}")
+    encoderblock_layernorm1_bias = state_dict.pop(f"img/Transformer/encoderblock/LayerNorm_1/bias{suffix}")
+
+    encoderblock_mlp_dense0_kernel= state_dict.pop(f"img/Transformer/encoderblock/MlpBlock_0/Dense_0/kernel{suffix}")
+    encoderblock_mlp_dense0_bias= state_dict.pop(f"img/Transformer/encoderblock/MlpBlock_0/Dense_0/bias{suffix}")
+    encoderblock_mlp_dense1_kernel= state_dict.pop(f"img/Transformer/encoderblock/MlpBlock_0/Dense_1/kernel{suffix}")
+    encoderblock_mlp_dense1_bias= state_dict.pop(f"img/Transformer/encoderblock/MlpBlock_0/Dense_1/bias{suffix}")
+
+    encoderblock_attention_0_key_kernel = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/key/kernel{suffix}")
+    encoderblock_attention_0_key_bias = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/key/bias{suffix}")
+    encoderblock_attention_0_value_kernel = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/value/kernel{suffix}")
+    encoderblock_attention_0_value_bias = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/value/bias{suffix}")
+    encoderblock_attention_0_query_kernel = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/query/kernel{suffix}")
+    encoderblock_attention_0_query_bias = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/query/bias{suffix}")
+    encoderblock_attention_0_out_kernel = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/out/kernel{suffix}")
+    encoderblock_attention_0_out_bias = state_dict.pop(f"img/Transformer/encoderblock/MultiHeadDotProductAttention_0/out/bias{suffix}")
+
+    for i in range(config.vision_config.num_hidden_layers):
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.layer_norm1.weight"] = encoderblock_layernorm0_scale[i].transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.layer_norm1.bias"] = encoderblock_layernorm0_bias[i]
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.layer_norm2.weight"] = encoderblock_layernorm1_scale[i].transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.layer_norm2.bias"] = encoderblock_layernorm1_bias[i]
+
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.mlp.fc1.weight"] = encoderblock_mlp_dense0_kernel[i].transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.mlp.fc1.bias"] = encoderblock_mlp_dense0_bias[i]
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.mlp.fc2.weight"] = encoderblock_mlp_dense1_kernel[i].transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.mlp.fc2.bias"] = encoderblock_mlp_dense1_bias[i]
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.k_proj.weight"] = encoderblock_attention_0_key_kernel[i].reshape(-1, config.vision_config.hidden_size).transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.k_proj.bias"] = encoderblock_attention_0_key_bias[i].reshape(-1, config.vision_config.hidden_size).reshape(-1)
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.v_proj.weight"] = encoderblock_attention_0_value_kernel[i].reshape(-1, config.vision_config.hidden_size).transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.v_proj.bias"] = encoderblock_attention_0_value_bias[i].reshape(-1, config.vision_config.hidden_size).reshape(-1)
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.q_proj.weight"] = encoderblock_attention_0_query_kernel[i].reshape(-1, config.vision_config.hidden_size).transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.q_proj.bias"] = encoderblock_attention_0_query_bias[i].reshape(-1, config.vision_config.hidden_size).reshape(-1)
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.out_proj.weight"] = encoderblock_attention_0_out_kernel[i].reshape(-1, config.vision_config.hidden_size).transpose()
+        state_dict[f"paligemma.vision_tower.vision_model.encoder.layers.{i}.self_attn.out_proj.bias"] = encoderblock_attention_0_out_bias[i].reshape(-1, config.vision_config.hidden_size).reshape(-1)
+
+    state_dict["paligemma.vision_tower.vision_model.post_layernorm.weight"] = state_dict.pop(f"img/Transformer/encoder_norm/scale{suffix}").transpose()
+    state_dict["paligemma.vision_tower.vision_model.post_layernorm.bias"] = state_dict.pop(f"img/Transformer/encoder_norm/bias{suffix}")
+
+    # multimodal projector
+
+    state_dict['paligemma.multi_modal_projector.linear.weight'] = state_dict.pop(f"img/head/kernel{suffix}").transpose()
+    state_dict['paligemma.multi_modal_projector.linear.bias'] = state_dict.pop(f"img/head/bias{suffix}")
+
+    # text decoder (gemma)
+    embedding_vector = state_dict.pop(f"llm/embedder/input_embedding{suffix}")
+    state_dict["paligemma.language_model.model.embed_tokens.weight"] = embedding_vector
+
+    # pop the einsum attention + mlp representations. There are 18 layers in gemma-2b.
+
+    llm_attention_attn_vec_einsum = state_dict.pop(f"llm/layers/attn/attn_vec_einsum/w{suffix}")
+    llm_attention_kv_einsum = state_dict.pop(f"llm/layers/attn/kv_einsum/w{suffix}")
+    llm_attention_q_einsum = state_dict.pop(f"llm/layers/attn/q_einsum/w{suffix}")
+
+    llm_mlp_gating_einsum = state_dict.pop(f"llm/layers/mlp/gating_einsum{suffix}")
+    llm_mlp_linear = state_dict.pop(f"llm/layers/mlp/linear{suffix}")
+    # TODO verify correctness of layer norm loading
+
+    llm_input_layernorm = state_dict.pop(f"llm/layers/pre_attention_norm/scale{suffix}")
+    llm_post_attention_layernorm = state_dict.pop(f"llm/layers/pre_ffw_norm/scale{suffix}")
+
+    for i in range(config.text_config.num_hidden_layers):
+        # llm_attention_q_einsum[i].shape = (8, 2048, 256)
+        q_proj_weight_reshaped = llm_attention_q_einsum[i].transpose(0, 2, 1).reshape(config.text_config.num_attention_heads * config.text_config.head_dim, config.text_config.hidden_size)
+
+        state_dict[f"paligemma.language_model.model.layers.{i}.self_attn.q_proj.weight"] = q_proj_weight_reshaped
+
+        # llm_attention_kv_einsum[i, 0, 0].shape = (2048, 256)
+        k_proj_weight_reshaped = llm_attention_kv_einsum[i, 0, 0].transpose()
+        state_dict[f"paligemma.language_model.model.layers.{i}.self_attn.k_proj.weight"] = k_proj_weight_reshaped
+        # llm_attention_kv_einsum[i, 1, 0].shape = (2048, 256)
+        v_proj_weight_reshaped = llm_attention_kv_einsum[i, 1, 0].transpose()
+        state_dict[f"paligemma.language_model.model.layers.{i}.self_attn.v_proj.weight"] = v_proj_weight_reshaped
+
+        # output projection.
+
+        # llm_attention_attn_vec_einsum[i].shape = (8, 256, 2048)
+        o_proj_weight_reshaped = llm_attention_attn_vec_einsum[i].transpose(2, 0, 1).reshape(config.text_config.num_attention_heads * config.text_config.head_dim, config.text_config.hidden_size)
+
+        state_dict[f"paligemma.language_model.model.layers.{i}.self_attn.o_proj.weight"] = o_proj_weight_reshaped
+        # mlp layers
+        gate_proj_weight = llm_mlp_gating_einsum[i, 0]
+        state_dict[f"paligemma.language_model.model.layers.{i}.mlp.gate_proj.weight"] = gate_proj_weight.transpose()
+        up_proj_weight = llm_mlp_gating_einsum[i, 1]
+        state_dict[f"paligemma.language_model.model.layers.{i}.mlp.up_proj.weight"] = up_proj_weight.transpose()
+        state_dict[f"paligemma.language_model.model.layers.{i}.mlp.down_proj.weight"] = llm_mlp_linear[i].transpose()
+        state_dict[f"paligemma.language_model.model.layers.{i}.input_layernorm.weight"] = llm_input_layernorm[i]
+        state_dict[f"paligemma.language_model.model.layers.{i}.post_attention_layernorm.weight"] = llm_post_attention_layernorm[i]
+
+    state_dict["paligemma.language_model.model.norm.weight"] = state_dict.pop(f"llm/final_norm/scale{suffix}")
+    state_dict["paligemma.language_model.lm_head.weight"] = embedding_vector # weights are tied.
+
+    # fmt: on
+    expert_dict = {}
+    final_state_dict = {}
+    for key, value in state_dict.items():
+        if key not in [
+            f"llm/final_norm_1/scale{suffix}",
+            f"llm/layers/attn/attn_vec_einsum_1/w{suffix}",
+            f"llm/layers/attn/kv_einsum_1/w{suffix}",
+            f"llm/layers/attn/q_einsum_1/w{suffix}",
+            f"llm/layers/mlp_1/gating_einsum{suffix}",
+            f"llm/layers/mlp_1/linear{suffix}",
+            f"llm/layers/pre_attention_norm_1/scale{suffix}",
+            f"llm/layers/pre_ffw_norm_1/scale{suffix}",
+        ]:
+            final_state_dict[key] = torch.from_numpy(value)
+        else:
+            expert_dict[key] = value
+
+    return final_state_dict, expert_dict
+
+
+def slice_gemma_state_dict(state_dict, config, num_expert=1):
+    # fmt: off
+    # text decoder (gemma)
+    # no embedding vector, the expert just has the decoder layers
+
+    embedding_vector = torch.zeros([config.vocab_size, config.hidden_size])
+    state_dict["gemma_expert.model.embed_tokens.weight"] = embedding_vector
+
+    # pop the einsum attention + mlp representations. There are 18 layers in gemma-2b.
+
+    suffix = "/value" if f"llm/layers/attn/attn_vec_einsum_{num_expert}/w/value" in state_dict else ""
+
+    llm_attention_attn_vec_einsum = state_dict.pop(f"llm/layers/attn/attn_vec_einsum_{num_expert}/w{suffix}")
+    llm_attention_kv_einsum = state_dict.pop(f"llm/layers/attn/kv_einsum_{num_expert}/w{suffix}")
+    llm_attention_q_einsum = state_dict.pop(f"llm/layers/attn/q_einsum_{num_expert}/w{suffix}")
+
+    llm_mlp_gating_einsum = state_dict.pop(f"llm/layers/mlp_{num_expert}/gating_einsum{suffix}")
+    llm_mlp_linear = state_dict.pop(f"llm/layers/mlp_{num_expert}/linear{suffix}")
+    # TODO verify correctness of layer norm loading
+
+    llm_input_layernorm = state_dict.pop(f"llm/layers/pre_attention_norm_{num_expert}/scale{suffix}")
+    llm_post_attention_layernorm = state_dict.pop(f"llm/layers/pre_ffw_norm_{num_expert}/scale{suffix}")
+
+    for i in range(config.num_hidden_layers):
+        q_proj_weight_reshaped = llm_attention_q_einsum[i].transpose(0, 2, 1).reshape(config.num_attention_heads * config.head_dim, config.hidden_size)
+
+        state_dict[f"gemma_expert.model.layers.{i}.self_attn.q_proj.weight"] = q_proj_weight_reshaped
+
+        k_proj_weight_reshaped = llm_attention_kv_einsum[i, 0, 0].transpose()
+        state_dict[f"gemma_expert.model.layers.{i}.self_attn.k_proj.weight"] = k_proj_weight_reshaped
+        v_proj_weight_reshaped = llm_attention_kv_einsum[i, 1, 0].transpose()
+        state_dict[f"gemma_expert.model.layers.{i}.self_attn.v_proj.weight"] = v_proj_weight_reshaped
+
+        # output projection.
+
+        # llm_attention_attn_vec_einsum[i].shape = (8, 256, 1024)
+        o_proj_weight_reshaped = llm_attention_attn_vec_einsum[i].reshape(config.num_attention_heads * config.head_dim, config.hidden_size).transpose(1,0)# .transpose(2, 0, 1).reshape(config.num_attention_heads * config.head_dim, config.hidden_size).transpose(1, 0)
+
+        state_dict[f"gemma_expert.model.layers.{i}.self_attn.o_proj.weight"] = o_proj_weight_reshaped
+        # mlp layers
+        gate_proj_weight = llm_mlp_gating_einsum[i, 0]
+        state_dict[f"gemma_expert.model.layers.{i}.mlp.gate_proj.weight"] = gate_proj_weight.transpose()
+        up_proj_weight = llm_mlp_gating_einsum[i, 1]
+        state_dict[f"gemma_expert.model.layers.{i}.mlp.up_proj.weight"] = up_proj_weight.transpose()
+        state_dict[f"gemma_expert.model.layers.{i}.mlp.down_proj.weight"] = llm_mlp_linear[i].transpose()
+        state_dict[f"gemma_expert.model.layers.{i}.input_layernorm.weight"] = llm_input_layernorm[i]
+        state_dict[f"gemma_expert.model.layers.{i}.post_attention_layernorm.weight"] = llm_post_attention_layernorm[i]
+
+    state_dict["gemma_expert.model.norm.weight"] = state_dict.pop(f"llm/final_norm_{num_expert}/scale{suffix}")
+    state_dict["gemma_expert.lm_head.weight"] = embedding_vector # weights are tied. (and zeros here)
+
+    # fmt: on
+    final_state_dict = {}
+    for key, value in state_dict.items():
+        if not isinstance(value, torch.Tensor):
+            final_state_dict[key] = torch.from_numpy(value)
+        else:
+            final_state_dict[key] = value
+    return final_state_dict
+
+
+def flatten_for_memory(tree, parent_key=""):
+    out = {}
+    for k, v in tree.items():
+        new_key = f"{parent_key}/{k}" if parent_key else k
+        if isinstance(v, dict):
+            out.update(flatten_for_memory(v, new_key))
+        else:
+            out[new_key] = np.array(v)  # Ensure conversion to np.array for consistency
+    return out
+
+
+def flatten_for_npz(tree, parent_key=""):
+    out = {}
+    for k, v in tree.items():
+        new_key = f"{parent_key}/{k}" if parent_key else k
+        if isinstance(v, dict):
+            out.update(flatten_for_npz(v, new_key))
+        else:
+            # bf16/f32 here?
+            out[new_key] = np.array(v)
+    return out
+
+
+def slice_initial_orbax_checkpoint(checkpoint_dir: str):
+    params_path = pathlib.Path(checkpoint_dir).resolve()
+    checkpointer = ocp.PyTreeCheckpointer()
+
+    metadata = checkpointer.metadata(params_path)
+    print("Metadata keys:", list(metadata.keys()))
+
+    params_name = "params"
+
+    item = {params_name: metadata[params_name]}
+    device = jax.local_devices()[0]  # Use the first local device
+    sharding = SingleDeviceSharding(device)
+    restored = checkpointer.restore(
+        params_path,
+        ocp.args.PyTreeRestore(
+            item=item,
+            restore_args=jax.tree_util.tree_map(
+                lambda _: ocp.ArrayRestoreArgs(
+                    restore_type=jax.Array,  # or np.ndarray, but bf16 is annoying about it
+                    sharding=sharding,
+                ),
+                item,
+            ),
+            transforms={},
+        ),
+    )
+    params = restored[params_name]
+
+    # get params for PaliGemma
+    pali_params = params["PaliGemma"]
+    del params["PaliGemma"]
+    pali_params_flat = flatten_for_npz(pali_params)
+    return {"paligemma_params": pali_params_flat, "projection_params": params}
+
+
+def update_keys_with_prefix(d: dict, prefix: str) -> dict:
+    """Update dictionary keys by adding a prefix."""
+    return {f"{prefix}{key}": value for key, value in d.items()}
+
+
+def convert_pi0_checkpoint(checkpoint_dir: str, precision: str, tokenizer_id: str, output_path: str):
+    # Break down orbax ckpts - they are in OCDBT
+    initial_params = slice_initial_orbax_checkpoint(checkpoint_dir=checkpoint_dir)
+    # process projection params
+    keys = [
+        "state_proj",
+        "action_in_proj",
+        "action_out_proj",
+        "action_time_mlp_in",
+        "action_time_mlp_out",
+    ]
+
+    projection_params = {}
+    for key in keys:
+        kernel_params = initial_params["projection_params"][key]["kernel"]
+        bias_params = initial_params["projection_params"][key]["bias"]
+        if isinstance(kernel_params, dict):
+            weight = kernel_params["value"]
+            bias = bias_params["value"]
+        else:
+            weight = kernel_params
+            bias = bias_params
+        projection_params[f"{key}.weight"] = torch.from_numpy(np.array(weight)).T
+        projection_params[f"{key}.bias"] = torch.from_numpy(np.array(bias))
+
+    # Process PaliGemma weights
+    paligemma_config = get_paligemma_config(precision)
+    paligemma_params, gemma_raw_dictionary = slice_paligemma_state_dict(
+        initial_params["paligemma_params"], paligemma_config
+    )
+
+    # Process Gemma  weights (at this stage they are unused)
+    gemma_config = get_gemma_config(precision)
+    gemma_params = slice_gemma_state_dict(gemma_raw_dictionary, config=gemma_config)
+
+    # Instantiate model from configs
+
+    if "pi0_aloha_sim" in checkpoint_dir:
+        pi0_config = PI0Config(
+            empty_cameras=2,
+            adapt_to_pi_aloha=True,
+            use_delta_joint_actions_aloha=False,
+        )
+    elif "pi0_aloha_towel" in checkpoint_dir:
+        pi0_config = PI0Config(
+            adapt_to_pi_aloha=True,
+            use_delta_joint_actions_aloha=True,
+        )
+    elif "pi0_base" in checkpoint_dir:
+        pi0_config = PI0Config(
+            empty_cameras=0,
+            adapt_to_pi_aloha=False,
+            use_delta_joint_actions_aloha=False,
+        )
+    else:
+        raise ValueError()
+
+    # gemma_config=gemma_config, paligemma_config=paligemma_config)
+    pi0_model = PI0Policy(pi0_config)
+
+    paligemma_params = update_keys_with_prefix(paligemma_params, "model.paligemma_with_expert.")
+    gemma_params = update_keys_with_prefix(gemma_params, "model.paligemma_with_expert.")
+    projection_params = update_keys_with_prefix(projection_params, "model.")
+
+    # load state dict
+    torch_dtype = PRECISIONS[precision]
+    pi0_model.load_state_dict({**paligemma_params, **gemma_params, **projection_params})
+    pi0_model = pi0_model.to(torch_dtype)
+    # pi0_tokenizer = AutoTokenizer.from_pretrained(tokenizer_id)
+
+    pi0_model.save_pretrained(output_path, safe_serialization=True)
+    # pi0_tokenizer.save_pretrained(output_path, dtype=torch_dtype)
+
+    # assert that model loads properly
+    del pi0_model
+    PI0Policy.from_pretrained(output_path)
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--checkpoint_dir",
+        default="/raid/pablo/.cache/openpi/openpi-assets/checkpoints/pi0_aloha_sim/params",
+        type=str,
+        help="Path to the ocdbt checkpoint",
+    )
+
+    parser.add_argument(
+        "--precision",
+        choices=["float32", "bfloat16", "float16"],
+        default="float32",
+        type=str,
+        help="Precision identifier for model conversion - should match the base checkpoint precision.",
+    )
+    # tokenizer is identical to paligemma, it appears
+
+    parser.add_argument(
+        "--tokenizer_hub_id",
+        default="google/paligemma-3b-pt-224",
+        type=str,
+        help="Hub path to the tokenizer to save",
+    )
+
+    parser.add_argument(
+        "--output_path",
+        required=True,
+        type=str,
+        help="Path to save converted weights to",
+    )
+
+    args = parser.parse_args()
+    convert_pi0_checkpoint(
+        checkpoint_dir=args.checkpoint_dir,
+        precision=args.precision,
+        tokenizer_id=args.tokenizer_hub_id,
+        output_path=args.output_path,
+    )

lerobot/common/policies/pi0/flex_attention.py

@@ -0,0 +1,127 @@
+import torch
+import torch.nn.functional as F  # noqa: N812
+from packaging.version import Version
+
+if Version(torch.__version__) > Version("2.5.0"):
+    # Ffex attention is only available from torch 2.5 onwards
+    from torch.nn.attention.flex_attention import (
+        _mask_mod_signature,
+        _round_up_to_multiple,
+        create_block_mask,
+        create_mask,
+        flex_attention,
+    )
+
+
+# @torch.compile(dynamic=False)
+def flex_attention_forward(
+    attention_mask: torch.Tensor,
+    batch_size: int,
+    head_dim: int,
+    query_states: torch.Tensor,
+    key_states: torch.Tensor,
+    value_states: torch.Tensor,
+    scaling=None,
+):
+    """
+    This is defined out of classes to make compile happy.
+    """
+
+    original_dtype = query_states.dtype
+    num_att_heads = 8
+    num_key_value_heads = 1
+    num_key_value_groups = num_att_heads // num_key_value_heads
+
+    key_states = key_states[:, :, :, None, :]
+    key_states = key_states.expand(
+        batch_size, key_states.shape[1], num_key_value_heads, num_key_value_groups, head_dim
+    )
+    key_states = key_states.reshape(
+        batch_size, key_states.shape[1], num_key_value_heads * num_key_value_groups, head_dim
+    )
+
+    value_states = value_states[:, :, :, None, :]
+    value_states = value_states.expand(
+        batch_size, value_states.shape[1], num_key_value_heads, num_key_value_groups, head_dim
+    )
+    value_states = value_states.reshape(
+        batch_size, value_states.shape[1], num_key_value_heads * num_key_value_groups, head_dim
+    )
+
+    query_states = query_states.transpose(1, 2)
+    key_states = key_states.transpose(1, 2)
+    value_states = value_states.transpose(1, 2)
+
+    query_states = query_states.to(torch.float32)
+    key_states = key_states.to(torch.float32)
+    value_states = value_states.to(torch.float32)
+
+    causal_mask = attention_mask
+    if causal_mask is not None:
+        causal_mask = causal_mask[:, None, :, : key_states.shape[2]]
+
+        if causal_mask.shape[1] == 1 and query_states.shape[1] > 1:
+            causal_mask = causal_mask.expand(-1, query_states.shape[1], -1, -1)
+
+    def precomputed_mask_factory(precomputed_mask: torch.Tensor) -> _mask_mod_signature:
+        def mask_mod(b, h, q_idx, kv_idx):
+            # Danger zone: if b,h,q_idx,kv_idx exceed the shape, device-side assert occurs.
+            return precomputed_mask[b][h][q_idx][kv_idx]
+
+        return mask_mod
+
+    b_mask, h_mask, q_len, kv_len = causal_mask.shape  # The shape of your mask
+
+    block_size = 128
+    q_len_rounded = _round_up_to_multiple(q_len, block_size)
+    kv_len_rounded = _round_up_to_multiple(kv_len, block_size)
+
+    # *CRITICAL* we do need to expand here, else we get a CUDA index error
+
+    pad_q = q_len_rounded - q_len
+    pad_k = kv_len_rounded - kv_len
+
+    padded_causal_mask = F.pad(causal_mask, (0, pad_k, 0, pad_q), value=0.0)
+    mask_mod_fn_orig = precomputed_mask_factory(padded_causal_mask)
+
+    mask_4d = create_mask(
+        mod_fn=mask_mod_fn_orig,
+        B=b_mask,
+        H=h_mask,
+        Q_LEN=q_len_rounded,
+        KV_LEN=kv_len_rounded,
+        device=causal_mask.device,
+        _compile=False,
+    )
+
+    mask_mod_fn_padded = precomputed_mask_factory(mask_4d)
+    block_mask = create_block_mask(
+        mask_mod=mask_mod_fn_padded,
+        B=b_mask,
+        H=h_mask,
+        Q_LEN=q_len_rounded,
+        KV_LEN=kv_len_rounded,
+        BLOCK_SIZE=block_size,
+        device=causal_mask.device,
+        _compile=False,
+    )
+
+    #  mask is applied inside the kernel, ideally more efficiently than score_mod.
+    attn_output, attention_weights = flex_attention(
+        query_states,
+        key_states,
+        value_states,
+        block_mask=block_mask,
+        enable_gqa=True,  # because we shaped query/key states for GQA
+        scale=head_dim**-0.5 if scaling is None else scaling,
+        return_lse=True,
+    )
+
+    attn_output = attn_output.to(dtype=original_dtype)
+    attn_output = attn_output.transpose(1, 2).contiguous()  # [B, Q_LEN, H, head_dim]
+    attn_output = attn_output.reshape(
+        batch_size,
+        -1,
+        attn_output.shape[2] * attn_output.shape[3],  # merges [H, head_dim]
+    )
+    return attn_output

lerobot/common/policies/pi0/modeling_pi0.py

@@ -0,0 +1,732 @@
+#!/usr/bin/env python
+
+# Copyright 2025 Physical Intelligence and 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.
+
+"""
+π0: A Vision-Language-Action Flow Model for General Robot Control
+
+[Paper](https://www.physicalintelligence.company/download/pi0.pdf)
+[Jax code](https://github.com/Physical-Intelligence/openpi)
+
+Designed by Physical Intelligence. Ported from Jax by Hugging Face.
+
+Install pi0 extra dependencies:
+```bash
+pip install -e ".[pi0]"
+```
+
+Example of finetuning the pi0 pretrained model (`pi0_base` in `openpi`):
+```bash
+python lerobot/scripts/train.py \
+--policy.path=lerobot/pi0 \
+--dataset.repo_id=danaaubakirova/koch_test
+```
+
+Example of finetuning the pi0 neural network with PaliGemma and expert Gemma
+pretrained with VLM default parameters before pi0 finetuning:
+```bash
+python lerobot/scripts/train.py \
+--policy.type=pi0 \
+--dataset.repo_id=danaaubakirova/koch_test
+```
+
+Example of using the pi0 pretrained model outside LeRobot training framework:
+```python
+policy = Pi0Policy.from_pretrained("lerobot/pi0")
+```
+
+"""
+
+import math
+from collections import deque
+
+import torch
+import torch.nn.functional as F  # noqa: N812
+from torch import Tensor, nn
+from transformers import AutoTokenizer
+
+from lerobot.common.constants import ACTION, OBS_ROBOT
+from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
+from lerobot.common.policies.pi0.paligemma_with_expert import (
+    PaliGemmaWithExpertConfig,
+    PaliGemmaWithExpertModel,
+)
+from lerobot.common.policies.pretrained import PreTrainedPolicy
+from lerobot.common.utils.utils import get_safe_dtype
+
+
+def create_sinusoidal_pos_embedding(
+    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]
+    pos_emb = torch.cat([torch.sin(sin_input), torch.cos(sin_input)], dim=1)
+    return pos_emb
+
+
+def sample_beta(alpha, beta, bsize, device):
+    gamma1 = torch.empty((bsize,), device=device).uniform_(0, 1).pow(1 / alpha)
+    gamma2 = torch.empty((bsize,), device=device).uniform_(0, 1).pow(1 / beta)
+    return gamma1 / (gamma1 + gamma2)
+
+
+def make_att_2d_masks(pad_masks, att_masks):
+    """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]
+    att_2d_masks = att_2d_masks & pad_2d_masks
+    return att_2d_masks
+
+
+def resize_with_pad(img, width, height, pad_value=-1):
+    # assume no-op when width height fits already
+    if img.ndim != 4:
+        raise ValueError(f"(b,c,h,w) expected, but {img.shape}")
+
+    cur_height, cur_width = img.shape[2:]
+
+    ratio = max(cur_width / width, cur_height / height)
+    resized_height = int(cur_height / ratio)
+    resized_width = int(cur_width / ratio)
+    resized_img = F.interpolate(
+        img, size=(resized_height, resized_width), mode="bilinear", align_corners=False
+    )
+
+    pad_height = max(0, int(height - resized_height))
+    pad_width = max(0, int(width - resized_width))
+
+    # pad on left and top of image
+    padded_img = F.pad(resized_img, (pad_width, 0, pad_height, 0), value=pad_value)
+    return padded_img
+
+
+def pad_vector(vector, new_dim):
+    """Can be (batch_size x sequence_length x features_dimension)
+    or (batch_size x features_dimension)
+    """
+    if vector.shape[-1] == new_dim:
+        return vector
+    shape = list(vector.shape)
+    current_dim = shape[-1]
+    shape[-1] = new_dim
+    new_vector = torch.zeros(*shape, dtype=vector.dtype, device=vector.device)
+    new_vector[..., :current_dim] = vector
+    return new_vector
+
+
+def normalize(x, min_val, max_val):
+    return (x - min_val) / (max_val - min_val)
+
+
+def unnormalize(x, min_val, max_val):
+    return x * (max_val - min_val) + min_val
+
+
+def safe_arcsin(value):
+    # This ensures that the input stays within
+    # [−1,1] to avoid invalid values for arcsin
+    return torch.arcsin(torch.clamp(value, -1.0, 1.0))
+
+
+def aloha_gripper_to_angular(value):
+    # Aloha transforms the gripper positions into a linear space. The following code
+    # reverses this transformation to be consistent with pi0 which is pretrained in
+    # angular space.
+    #
+    # These values are coming from the Aloha code:
+    # PUPPET_GRIPPER_POSITION_OPEN, PUPPET_GRIPPER_POSITION_CLOSED
+    value = unnormalize(value, min_val=0.01844, max_val=0.05800)
+
+    # This is the inverse of the angular to linear transformation inside the Interbotix code.
+    def linear_to_radian(linear_position, arm_length, horn_radius):
+        value = (horn_radius**2 + linear_position**2 - arm_length**2) / (2 * horn_radius * linear_position)
+        return safe_arcsin(value)
+
+    # The constants are taken from the Interbotix code.
+    value = linear_to_radian(value, arm_length=0.036, horn_radius=0.022)
+
+    # Normalize to [0, 1].
+    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
+    return normalize(value, min_val=0.4, max_val=1.5)
+
+
+def aloha_gripper_from_angular(value):
+    # Convert from the gripper position used by pi0 to the gripper position that is used by Aloha.
+    # Note that the units are still angular but the range is different.
+
+    # The values 0.4 and 1.5 were measured on an actual Trossen robot.
+    value = unnormalize(value, min_val=0.4, max_val=1.5)
+
+    # These values are coming from the Aloha code:
+    # PUPPET_GRIPPER_JOINT_OPEN, PUPPET_GRIPPER_JOINT_CLOSE
+    return normalize(value, min_val=-0.6213, max_val=1.4910)
+
+
+def aloha_gripper_from_angular_inv(value):
+    # Directly inverts the gripper_from_angular function.
+    value = unnormalize(value, min_val=-0.6213, max_val=1.4910)
+    return normalize(value, min_val=0.4, max_val=1.5)
+
+
+class PI0Policy(PreTrainedPolicy):
+    """Wrapper class around PI0FlowMatching model to train and run inference within LeRobot."""
+
+    config_class = PI0Config
+    name = "pi0"
+
+    def __init__(
+        self,
+        config: PI0Config,
+        dataset_stats: dict[str, dict[str, Tensor]] | None = None,
+    ):
+        """
+        Args:
+            config: Policy configuration class instance or None, in which case the default instantiation of
+                    the configuration class is used.
+            dataset_stats: Dataset statistics to be used for normalization. If not passed here, it is expected
+                that they will be passed with a call to `load_state_dict` before the policy is used.
+        """
+
+        super().__init__(config)
+        config.validate_features()
+        self.config = config
+        self.normalize_inputs = Normalize(config.input_features, config.normalization_mapping, dataset_stats)
+        self.normalize_targets = Normalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+        self.unnormalize_outputs = Unnormalize(
+            config.output_features, config.normalization_mapping, dataset_stats
+        )
+
+        self.language_tokenizer = AutoTokenizer.from_pretrained("google/paligemma-3b-pt-224")
+        self.model = PI0FlowMatching(config)
+
+        self.reset()
+
+    def reset(self):
+        """This should be called whenever the environment is reset."""
+        self._action_queue = deque([], maxlen=self.config.n_action_steps)
+
+    def get_optim_params(self) -> dict:
+        return self.parameters()
+
+    @torch.no_grad
+    def select_action(self, batch: dict[str, Tensor], noise: Tensor | None = None) -> Tensor:
+        """Select a single action given environment observations.
+
+        This method wraps `select_actions` in order to return one action at a time for execution in the
+        environment. It works by managing the actions in a queue and only calling `select_actions` when the
+        queue is empty.
+        """
+        self.eval()
+
+        if self.config.adapt_to_pi_aloha:
+            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
+
+        batch = self.normalize_inputs(batch)
+
+        # Action queue logic for n_action_steps > 1. When the action_queue is depleted, populate it by
+        # querying the policy.
+        if len(self._action_queue) == 0:
+            images, img_masks = self.prepare_images(batch)
+            state = self.prepare_state(batch)
+            lang_tokens, lang_masks = self.prepare_language(batch)
+
+            actions = self.model.sample_actions(
+                images, img_masks, lang_tokens, lang_masks, state, noise=noise
+            )
+
+            # Unpad actions
+            original_action_dim = self.config.action_feature.shape[0]
+            actions = actions[:, :, :original_action_dim]
+
+            actions = self.unnormalize_outputs({"action": actions})["action"]
+
+            if self.config.adapt_to_pi_aloha:
+                actions = self._pi_aloha_encode_actions(actions)
+
+            # `self.model.forward` returns a (batch_size, n_action_steps, action_dim) tensor, but the queue
+            # effectively has shape (n_action_steps, batch_size, *), hence the transpose.
+            self._action_queue.extend(actions.transpose(0, 1))
+        return self._action_queue.popleft()
+
+    def forward(self, batch: dict[str, Tensor], noise=None, time=None) -> dict[str, Tensor]:
+        """Do a full training forward pass to compute the loss"""
+        if self.config.adapt_to_pi_aloha:
+            batch[OBS_ROBOT] = self._pi_aloha_decode_state(batch[OBS_ROBOT])
+            batch[ACTION] = self._pi_aloha_encode_actions_inv(batch[ACTION])
+
+        batch = self.normalize_inputs(batch)
+        batch = self.normalize_targets(batch)
+
+        images, img_masks = self.prepare_images(batch)
+        state = self.prepare_state(batch)
+        lang_tokens, lang_masks = self.prepare_language(batch)
+        actions = self.prepare_action(batch)
+        actions_is_pad = batch.get("actions_id_pad")
+
+        loss_dict = {}
+        losses = self.model.forward(images, img_masks, lang_tokens, lang_masks, state, actions, noise, time)
+        loss_dict["losses_after_forward"] = losses.clone()
+
+        if actions_is_pad is not None:
+            in_episode_bound = ~actions_is_pad
+            losses = losses * in_episode_bound.unsqueeze(-1)
+            loss_dict["losses_after_in_ep_bound"] = losses.clone()
+
+        # Remove padding
+        losses = losses[:, :, : self.config.max_action_dim]
+        loss_dict["losses_after_rm_padding"] = losses.clone()
+
+        loss = losses.mean()
+        # For backward pass
+        loss_dict["loss"] = loss
+        # For logging
+        loss_dict["l2_loss"] = loss.item()
+        return loss_dict
+
+    def prepare_images(self, batch):
+        """Apply Pi0 preprocessing to the images, like resizing to 224x224 and padding to keep aspect ratio, and
+        convert pixel range from [0.0, 1.0] to [-1.0, 1.0] as requested by SigLIP.
+        """
+        images = []
+        img_masks = []
+
+        present_img_keys = [key for key in self.config.image_features if key in batch]
+        missing_img_keys = [key for key in self.config.image_features if key not in batch]
+
+        if len(present_img_keys) == 0:
+            raise ValueError(
+                f"All image features are missing from the batch. At least one expected. (batch: {batch.keys()}) (image_features:{self.config.image_features})"
+            )
+
+        # Preprocess image features present in the batch
+        for key in present_img_keys:
+            img = batch[key]
+
+            if self.config.resize_imgs_with_padding is not None:
+                img = resize_with_pad(img, *self.config.resize_imgs_with_padding, pad_value=0)
+
+            # Normalize from range [0,1] to [-1,1] as expacted by siglip
+            img = img * 2.0 - 1.0
+
+            bsize = img.shape[0]
+            device = img.device
+            mask = torch.ones(bsize, dtype=torch.bool, device=device)
+            images.append(img)
+            img_masks.append(mask)
+
+        # Create image features not present in the batch
+        # as fully 0 padded images.
+        for num_empty_cameras in range(len(missing_img_keys)):
+            if num_empty_cameras >= self.config.empty_cameras:
+                break
+            img = torch.ones_like(img) * -1
+            mask = torch.zeros_like(mask)
+            images.append(img)
+            img_masks.append(mask)
+
+        return images, img_masks
+
+    def prepare_language(self, batch) -> tuple[Tensor, Tensor]:
+        """Tokenize the text input"""
+        device = batch[OBS_ROBOT].device
+        tasks = batch["task"]
+
+        # PaliGemma prompt has to end with a new line
+        tasks = [task if task.endswith("\n") else f"{task}\n" for task in tasks]
+
+        tokenized_prompt = self.language_tokenizer.__call__(
+            tasks,
+            padding="max_length",
+            padding_side="right",
+            max_length=self.config.tokenizer_max_length,
+            return_tensors="pt",
+        )
+        lang_tokens = tokenized_prompt["input_ids"].to(device=device)
+        lang_masks = tokenized_prompt["attention_mask"].to(device=device, dtype=torch.bool)
+
+        return lang_tokens, lang_masks
+
+    def _pi_aloha_decode_state(self, state):
+        # Flip the joints.
+        for motor_idx in [1, 2, 8, 9]:
+            state[:, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            state[:, motor_idx] = aloha_gripper_to_angular(state[:, motor_idx])
+        return state
+
+    def _pi_aloha_encode_actions(self, actions):
+        # Flip the joints.
+        for motor_idx in [1, 2, 8, 9]:
+            actions[:, :, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            actions[:, :, motor_idx] = aloha_gripper_from_angular(actions[:, :, motor_idx])
+        return actions
+
+    def _pi_aloha_encode_actions_inv(self, actions):
+        # Flip the joints again.
+        for motor_idx in [1, 2, 8, 9]:
+            actions[:, :, motor_idx] *= -1
+        # Reverse the gripper transformation that is being applied by the Aloha runtime.
+        for motor_idx in [6, 13]:
+            actions[:, :, motor_idx] = aloha_gripper_from_angular_inv(actions[:, :, motor_idx])
+        return actions
+
+    def prepare_state(self, batch):
+        """Pad state"""
+        state = pad_vector(batch[OBS_ROBOT], self.config.max_state_dim)
+        return state
+
+    def prepare_action(self, batch):
+        """Pad action"""
+        actions = pad_vector(batch[ACTION], self.config.max_action_dim)
+        return actions
+
+
+class PI0FlowMatching(nn.Module):
+    """
+    π0: A Vision-Language-Action Flow Model for General Robot Control
+
+    [Paper](https://www.physicalintelligence.company/download/pi0.pdf)
+    [Jax code](https://github.com/Physical-Intelligence/openpi)
+
+    Designed by Physical Intelligence. Ported from Jax by Hugging Face.
+    ┌──────────────────────────────┐
+    │               actions        │
+    │               ▲              │
+    │              ┌┴─────┐        │
+    │  kv cache    │Gemma │        │
+    │  ┌──────────►│Expert│        │
+    │  │           │      │        │
+    │ ┌┴────────┐  │x 10  │        │
+    │ │         │  └▲──▲──┘        │
+    │ │PaliGemma│   │  │           │
+    │ │         │   │  robot state │
+    │ │         │   noise          │
+    │ └▲──▲─────┘                  │
+    │  │  │                        │
+    │  │  image(s)                 │
+    │  language tokens             │
+    └──────────────────────────────┘
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+
+        paligemma_with_export_config = PaliGemmaWithExpertConfig(
+            freeze_vision_encoder=self.config.freeze_vision_encoder,
+            train_expert_only=self.config.train_expert_only,
+            attention_implementation=self.config.attention_implementation,
+        )
+        self.paligemma_with_expert = PaliGemmaWithExpertModel(paligemma_with_export_config)
+
+        # Projections are float32
+        self.state_proj = nn.Linear(self.config.max_state_dim, self.config.proj_width)
+        self.action_in_proj = nn.Linear(self.config.max_action_dim, self.config.proj_width)
+        self.action_out_proj = nn.Linear(self.config.proj_width, self.config.max_action_dim)
+
+        self.action_time_mlp_in = nn.Linear(self.config.proj_width * 2, self.config.proj_width)
+        self.action_time_mlp_out = nn.Linear(self.config.proj_width, self.config.proj_width)
+
+        self.set_requires_grad()
+
+    def set_requires_grad(self):
+        for params in self.state_proj.parameters():
+            params.requires_grad = self.config.train_state_proj
+
+    def sample_noise(self, shape, device):
+        noise = torch.normal(
+            mean=0.0,
+            std=1.0,
+            size=shape,
+            dtype=torch.float32,
+            device=device,
+        )
+        return noise
+
+    def sample_time(self, bsize, device):
+        time_beta = sample_beta(1.5, 1.0, bsize, device)
+        time = time_beta * 0.999 + 0.001
+        return time.to(dtype=torch.float32, device=device)
+
+    def embed_prefix(
+        self, images, img_masks, lang_tokens, lang_masks
+    ) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """Embed images with SigLIP and language tokens with embedding layer to prepare
+        for PaliGemma transformer processing.
+        """
+        # TODO: avoid list in python and torch.cat ; prefer pre-allocation with torch.empty
+        embs = []
+        pad_masks = []
+        att_masks = []
+
+        # TODO: remove for loop
+        for (
+            img,
+            img_mask,
+        ) in zip(images, img_masks, strict=False):
+            img_emb = self.paligemma_with_expert.embed_image(img)
+            img_emb = img_emb.to(dtype=torch.bfloat16)
+
+            # Normalize image embeddings
+            img_emb_dim = img_emb.shape[-1]
+            img_emb = img_emb * torch.tensor(img_emb_dim**0.5, dtype=img_emb.dtype, device=img_emb.device)
+
+            bsize, num_img_embs = img_emb.shape[:2]
+            img_mask = img_mask[:, None].expand(bsize, num_img_embs)
+
+            embs.append(img_emb)
+            pad_masks.append(img_mask)
+
+            # Create attention masks so that image tokens attend to each other
+            att_masks += [0] * num_img_embs
+
+        lang_emb = self.paligemma_with_expert.embed_language_tokens(lang_tokens)
+
+        # Normalize language embeddings
+        lang_emb_dim = lang_emb.shape[-1]
+        lang_emb = lang_emb * math.sqrt(lang_emb_dim)
+
+        embs.append(lang_emb)
+        pad_masks.append(lang_masks)
+
+        # full attention between image and language inputs
+        num_lang_embs = lang_emb.shape[1]
+        att_masks += [0] * num_lang_embs
+
+        embs = torch.cat(embs, dim=1)
+        pad_masks = torch.cat(pad_masks, dim=1)
+        att_masks = torch.tensor(att_masks, dtype=torch.bool, device=pad_masks.device)
+        att_masks = att_masks[None, :].expand(bsize, len(att_masks))
+
+        return embs, pad_masks, att_masks
+
+    def embed_suffix(self, state, noisy_actions, timestep):
+        """Embed state, noisy_actions, timestep to prepare for Expert Gemma processing."""
+        embs = []
+        pad_masks = []
+        att_masks = []
+
+        # Embed state
+        state_emb = self.state_proj(state)
+        state_emb = state_emb.to(dtype=torch.bfloat16)
+        embs.append(state_emb[:, None, :])
+        bsize = state_emb.shape[0]
+        dtype = state_emb.dtype
+        device = state_emb.device
+
+        state_mask = torch.ones(bsize, 1, dtype=torch.bool, device=device)
+        pad_masks.append(state_mask)
+
+        # Set attention masks so that image and language inputs do not attend to state or actions
+        att_masks += [1]
+
+        # Embed timestep using sine-cosine positional encoding with sensitivity in the range [0, 1]
+        time_emb = create_sinusoidal_pos_embedding(
+            timestep, self.config.proj_width, min_period=4e-3, max_period=4.0, device=device
+        )
+        time_emb = time_emb.type(dtype=dtype)
+
+        # Fuse timestep + action information using an MLP
+        action_emb = self.action_in_proj(noisy_actions)
+
+        time_emb = time_emb[:, None, :].expand_as(action_emb)
+        action_time_emb = torch.cat([action_emb, time_emb], dim=2)
+
+        action_time_emb = self.action_time_mlp_in(action_time_emb)
+        action_time_emb = F.silu(action_time_emb)  # swish == silu
+        action_time_emb = self.action_time_mlp_out(action_time_emb)
+
+        # Add to input tokens
+        embs.append(action_time_emb)
+
+        bsize, action_time_dim = action_time_emb.shape[:2]
+        action_time_mask = torch.ones(bsize, action_time_dim, dtype=torch.bool, device=device)
+        pad_masks.append(action_time_mask)
+
+        # Set attention masks so that image, language and state inputs do not attend to action tokens
+        att_masks += [1] + ([0] * (self.config.n_action_steps - 1))
+
+        embs = torch.cat(embs, dim=1)
+        pad_masks = torch.cat(pad_masks, dim=1)
+        att_masks = torch.tensor(att_masks, dtype=embs.dtype, device=embs.device)
+        att_masks = att_masks[None, :].expand(bsize, len(att_masks))
+
+        return embs, pad_masks, att_masks
+
+    def forward(
+        self, images, img_masks, lang_tokens, lang_masks, state, actions, noise=None, time=None
+    ) -> Tensor:
+        """Do a full training forward pass and compute the loss (batch_size x num_steps x num_motors)"""
+        if noise is None:
+            noise = self.sample_noise(actions.shape, actions.device)
+
+        if time is None:
+            time = self.sample_time(actions.shape[0], actions.device)
+
+        time_expanded = time[:, None, None]
+        x_t = time_expanded * noise + (1 - time_expanded) * actions
+        u_t = noise - actions
+
+        prefix_embs, prefix_pad_masks, prefix_att_masks = self.embed_prefix(
+            images, img_masks, lang_tokens, lang_masks
+        )
+        suffix_embs, suffix_pad_masks, suffix_att_masks = self.embed_suffix(state, x_t, time)
+
+        pad_masks = torch.cat([prefix_pad_masks, suffix_pad_masks], dim=1)
+        att_masks = torch.cat([prefix_att_masks, suffix_att_masks], dim=1)
+
+        att_2d_masks = make_att_2d_masks(pad_masks, att_masks)
+        position_ids = torch.cumsum(pad_masks, dim=1) - 1
+
+        (_, suffix_out), _ = self.paligemma_with_expert.forward(
+            attention_mask=att_2d_masks,
+            position_ids=position_ids,
+            past_key_values=None,
+            inputs_embeds=[prefix_embs, suffix_embs],
+            use_cache=False,
+            fill_kv_cache=False,
+        )
+        suffix_out = suffix_out[:, -self.config.n_action_steps :]
+        # Original openpi code, upcast attention output
+        suffix_out = suffix_out.to(dtype=torch.float32)
+        v_t = self.action_out_proj(suffix_out)
+
+        losses = F.mse_loss(u_t, v_t, reduction="none")
+        return losses
+
+    def sample_actions(self, images, img_masks, lang_tokens, lang_masks, state, noise=None) -> Tensor:
+        """Do a full inference forward and compute the action (batch_size x num_steps x num_motors)"""
+        bsize = state.shape[0]
+        device = state.device
+
+        if noise is None:
+            actions_shape = (bsize, self.config.n_action_steps, self.config.max_action_dim)
+            noise = self.sample_noise(actions_shape, device)
+
+        prefix_embs, prefix_pad_masks, prefix_att_masks = self.embed_prefix(
+            images, img_masks, lang_tokens, lang_masks
+        )
+        prefix_att_2d_masks = make_att_2d_masks(prefix_pad_masks, prefix_att_masks)
+        prefix_position_ids = torch.cumsum(prefix_pad_masks, dim=1) - 1
+
+        # Compute image and language key value cache
+        _, past_key_values = self.paligemma_with_expert.forward(
+            attention_mask=prefix_att_2d_masks,
+            position_ids=prefix_position_ids,
+            past_key_values=None,
+            inputs_embeds=[prefix_embs, None],
+            use_cache=self.config.use_cache,
+            fill_kv_cache=True,
+        )
+
+        dt = -1.0 / self.config.num_steps
+        dt = torch.tensor(dt, dtype=torch.float32, device=device)
+
+        x_t = noise
+        time = torch.tensor(1.0, dtype=torch.float32, device=device)
+        while time >= -dt / 2:
+            expanded_time = time.expand(bsize)
+            v_t = self.denoise_step(
+                state,
+                prefix_pad_masks,
+                past_key_values,
+                x_t,
+                expanded_time,
+            )
+
+            # Euler step
+            x_t += dt * v_t
+            time += dt
+        return x_t
+
+    def denoise_step(
+        self,
+        state,
+        prefix_pad_masks,
+        past_key_values,
+        x_t,
+        timestep,
+    ):
+        """Apply one denoising step of the noise `x_t` at a given timestep."""
+        suffix_embs, suffix_pad_masks, suffix_att_masks = self.embed_suffix(state, x_t, timestep)
+
+        suffix_len = suffix_pad_masks.shape[1]
+        batch_size = prefix_pad_masks.shape[0]
+        prefix_len = prefix_pad_masks.shape[1]
+        prefix_pad_2d_masks = prefix_pad_masks[:, None, :].expand(batch_size, suffix_len, prefix_len)
+
+        suffix_att_2d_masks = make_att_2d_masks(suffix_pad_masks, suffix_att_masks)
+
+        full_att_2d_masks = torch.cat([prefix_pad_2d_masks, suffix_att_2d_masks], dim=2)
+
+        prefix_offsets = torch.sum(prefix_pad_masks, dim=-1)[:, None]
+        position_ids = prefix_offsets + torch.cumsum(suffix_pad_masks, dim=1) - 1
+
+        outputs_embeds, _ = self.paligemma_with_expert.forward(
+            attention_mask=full_att_2d_masks,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=[None, suffix_embs],
+            use_cache=self.config.use_cache,
+            fill_kv_cache=False,
+        )
+        suffix_out = outputs_embeds[1]
+        suffix_out = suffix_out[:, -self.config.n_action_steps :]
+        suffix_out = suffix_out.to(dtype=torch.float32)
+        v_t = self.action_out_proj(suffix_out)
+        return v_t

lerobot/common/policies/pi0/paligemma_with_expert.py

@@ -0,0 +1,403 @@
+from typing import List, Optional, Union
+
+import torch
+import torch.version
+from pytest import Cache
+from torch import nn
+from transformers import (
+    AutoConfig,
+    GemmaForCausalLM,
+    PaliGemmaForConditionalGeneration,
+    PretrainedConfig,
+    PreTrainedModel,
+)
+from transformers.models.auto import CONFIG_MAPPING
+
+from lerobot.common.policies.pi0.flex_attention import flex_attention_forward
+
+
+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)
+
+
+class PaliGemmaWithExpertConfig(PretrainedConfig):
+    model_type = "PaliGemmaWithExpertModel"
+    sub_configs = {"paligemma_config": AutoConfig, "gemma_expert_config": AutoConfig}
+
+    def __init__(
+        self,
+        paligemma_config: dict | None = None,
+        gemma_expert_config: dict | None = None,
+        freeze_vision_encoder: bool = True,
+        train_expert_only: bool = True,
+        attention_implementation: str = "eager",
+        **kwargs,
+    ):
+        self.freeze_vision_encoder = freeze_vision_encoder
+        self.train_expert_only = train_expert_only
+        self.attention_implementation = attention_implementation
+
+        if paligemma_config is None:
+            # Default config from Pi0
+            self.paligemma_config = CONFIG_MAPPING["paligemma"](
+                transformers_version="4.48.1",
+                _vocab_size=257152,
+                bos_token_id=2,
+                eos_token_id=1,
+                hidden_size=2048,
+                image_token_index=257152,
+                model_type="paligemma",
+                pad_token_id=0,
+                projection_dim=2048,
+                text_config={
+                    "hidden_activation": "gelu_pytorch_tanh",
+                    "hidden_size": 2048,
+                    "intermediate_size": 16384,
+                    "model_type": "gemma",
+                    "num_attention_heads": 8,
+                    "num_hidden_layers": 18,
+                    "num_image_tokens": 256,
+                    "num_key_value_heads": 1,
+                    "torch_dtype": "float32",
+                    "vocab_size": 257152,
+                },
+                vision_config={
+                    "hidden_size": 1152,
+                    "intermediate_size": 4304,
+                    "model_type": "siglip_vision_model",
+                    "num_attention_heads": 16,
+                    "num_hidden_layers": 27,
+                    "num_image_tokens": 256,
+                    "patch_size": 14,
+                    "projection_dim": 2048,
+                    "projector_hidden_act": "gelu_fast",
+                    "torch_dtype": "float32",
+                    "vision_use_head": False,
+                },
+            )
+        elif isinstance(self.paligemma_config, dict):
+            # Override Pi0 default config for PaliGemma
+            if "model_type" not in gemma_expert_config:
+                paligemma_config["model_type"] = "paligemma"
+
+            cfg_cls = CONFIG_MAPPING[paligemma_config["model_type"]]
+            self.paligemma_config = cfg_cls(**paligemma_config)
+
+        if gemma_expert_config is None:
+            # Default config from Pi0
+            self.gemma_expert_config = CONFIG_MAPPING["gemma"](
+                attention_bias=False,
+                attention_dropout=0.0,
+                bos_token_id=2,
+                eos_token_id=1,
+                head_dim=256,
+                hidden_act="gelu_pytorch_tanh",
+                hidden_activation="gelu_pytorch_tanh",
+                hidden_size=1024,
+                initializer_range=0.02,
+                intermediate_size=4096,
+                max_position_embeddings=8192,
+                model_type="gemma",
+                num_attention_heads=8,
+                num_hidden_layers=18,
+                num_key_value_heads=1,
+                pad_token_id=0,
+                rms_norm_eps=1e-06,
+                rope_theta=10000.0,
+                torch_dtype="float32",
+                transformers_version="4.48.1",
+                use_cache=True,
+                vocab_size=257152,
+            )
+        elif isinstance(self.gemma_expert_config, dict):
+            # Override Pi0 default config for Gemma Expert
+            if "model_type" not in gemma_expert_config:
+                gemma_expert_config["model_type"] = "gemma"
+
+            cfg_cls = CONFIG_MAPPING[paligemma_config["model_type"]]
+            self.gemma_expert_config = cfg_cls(**gemma_expert_config)
+
+        super().__init__(**kwargs)
+
+    def __post_init__(self):
+        super().__post_init__()
+        if self.train_expert_only and not self.freeze_vision_encoder:
+            raise ValueError(
+                "You set `freeze_vision_encoder=False` and `train_expert_only=True` which are not compatible."
+            )
+
+        if self.attention_implementation not in ["eager", "fa2", "flex"]:
+            raise ValueError(
+                f"Wrong value provided for `attention_implementation` ({self.attention_implementation}). Expected 'eager', 'fa2' or 'flex'."
+            )
+
+
+class PaliGemmaWithExpertModel(PreTrainedModel):
+    config_class = PaliGemmaWithExpertConfig
+
+    def __init__(self, config: PaliGemmaWithExpertConfig):
+        super().__init__(config=config)
+        self.config = config
+        self.paligemma = PaliGemmaForConditionalGeneration(config=config.paligemma_config)
+        self.gemma_expert = GemmaForCausalLM(config=config.gemma_expert_config)
+        # Remove unused embed_tokens
+        self.gemma_expert.model.embed_tokens = None
+
+        self.to_bfloat16_like_physical_intelligence()
+        self.set_requires_grad()
+
+    def set_requires_grad(self):
+        if self.config.freeze_vision_encoder:
+            self.paligemma.vision_tower.eval()
+            for params in self.paligemma.vision_tower.parameters():
+                params.requires_grad = False
+
+        if self.config.train_expert_only:
+            self.paligemma.eval()
+            for params in self.paligemma.parameters():
+                params.requires_grad = False
+
+    def train(self, mode: bool = True):
+        super().train(mode)
+
+        if self.config.freeze_vision_encoder:
+            self.paligemma.vision_tower.eval()
+
+        if self.config.train_expert_only:
+            self.paligemma.eval()
+
+    def to_bfloat16_like_physical_intelligence(self):
+        self.paligemma = self.paligemma.to(dtype=torch.bfloat16)
+
+        params_to_change_dtype = [
+            "language_model.model.layers",
+            "gemma_expert.model.layers",
+            "vision_tower",
+            "multi_modal",
+        ]
+        for name, param in self.named_parameters():
+            if any(selector in name for selector in params_to_change_dtype):
+                param.data = param.data.to(dtype=torch.bfloat16)
+
+    def embed_image(self, image: torch.Tensor):
+        return self.paligemma.get_image_features(image)
+
+    def embed_language_tokens(self, tokens: torch.Tensor):
+        return self.paligemma.language_model.model.embed_tokens(tokens)
+
+    # TODO: break down this huge forward into modules or functions
+    def forward(
+        self,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[Union[List[torch.FloatTensor], Cache]] = None,
+        inputs_embeds: List[torch.FloatTensor] = None,
+        use_cache: Optional[bool] = None,
+        fill_kv_cache: Optional[bool] = None,
+    ):
+        models = [self.paligemma.language_model.model, self.gemma_expert.model]
+
+        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.paligemma.config.text_config.num_hidden_layers
+        head_dim = self.paligemma.config.text_config.head_dim
+        for layer_idx in range(num_layers):
+            query_states = []
+            key_states = []
+            value_states = []
+            for i, hidden_states in enumerate(inputs_embeds):
+                if hidden_states is None:
+                    continue
+                layer = models[i].layers[layer_idx]
+                # normalizer = torch.tensor(models[i].config.hidden_size**0.5, dtype=hidden_states.dtype)
+                # hidden_states = hidden_states * normalizer
+                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=torch.bfloat16)
+                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)
+
+            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
+            )
+            att_output = att_output.to(dtype=torch.bfloat16)
+
+            # first part of att_output is prefix (up to sequence length, [:, 0:prefix_seq_len])
+            outputs_embeds = []
+            start = 0
+            for i, hidden_states in enumerate(inputs_embeds):
+                layer = models[i].layers[layer_idx]
+
+                if hidden_states is not None:
+                    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)
+                    out_emb = layer.self_attn.o_proj(att_output[:, start:end])
+
+                    # TODO: first dropout (by default 0.0)
+
+                    # first residual
+                    out_emb += hidden_states
+                    after_first_residual = out_emb.clone()
+
+                    out_emb = layer.post_attention_layernorm(out_emb)
+                    out_emb = layer.mlp(out_emb)
+
+                    # TODO: second dropout (by default 0.0)
+
+                    # second residual
+                    out_emb += after_first_residual
+
+                    outputs_embeds.append(out_emb)
+
+                    start = end
+                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):
+        if self.config.attention_implementation == "fa2":
+            attention_interface = self.flash_attention_forward
+        elif self.config.attention_implementation == "flex":
+            attention_interface = flex_attention_forward
+        else:
+            attention_interface = self.eager_attention_forward
+        return attention_interface
+
+    def flash_attention_forward(
+        self, attention_mask, batch_size, head_dim, query_states, key_states, value_states
+    ):
+        raise NotImplementedError("FA2 is not implemented (yet)")
+
+    def eager_attention_forward(
+        self, attention_mask, batch_size, head_dim, query_states, key_states, value_states
+    ):
+        num_att_heads = self.config.paligemma_config.text_config.num_attention_heads
+        num_key_value_heads = self.config.paligemma_config.text_config.num_key_value_heads
+        num_key_value_groups = num_att_heads // num_key_value_heads
+
+        # query_states: batch_size, sequence_length, num_att_head, head_dim
+        # key_states: batch_size, sequence_length, num_key_value_head, head_dim
+        # value_states: batch_size, sequence_length, num_key_value_head, head_dim
+        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
+        big_neg = -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)
+
+        # probs: batch_size, num_key_value_head, num_att_head, sequence_length, sequence_length
+        # value_states: batch_size, sequence_length, num_att_heads, head_dim
+
+        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

lerobot/common/policies/tdmpc/modeling_tdmpc.py

@@ -319,7 +319,7 @@ class TDMPCPolicy(PreTrainedPolicy):
 
         # (b, t) -> (t, b)
         for key in batch:
-            if batch[key].ndim > 1:
+            if isinstance(batch[key], torch.Tensor) and batch[key].ndim > 1:
                 batch[key] = batch[key].transpose(1, 0)
 
         action = batch["action"]  # (t, b, action_dim)
@@ -502,7 +502,7 @@ class TDMPCPolicy(PreTrainedPolicy):
 
         # Undo (b, t) -> (t, b).
         for key in batch:
-            if batch[key].ndim > 1:
+            if isinstance(batch[key], torch.Tensor) and batch[key].ndim > 1:
                 batch[key] = batch[key].transpose(1, 0)
 
         return info

lerobot/common/utils/utils.py

@@ -74,6 +74,18 @@ def get_safe_torch_device(try_device: str, log: bool = False) -> torch.device:
     return device
 
 
+def get_safe_dtype(dtype: torch.dtype, device: str | torch.device):
+    """
+    mps is currently not compatible with float64
+    """
+    if isinstance(device, torch.device):
+        device = device.type
+    if device == "mps" and dtype == torch.float64:
+        return torch.float32
+    else:
+        return dtype
+
+
 def is_torch_device_available(try_device: str) -> bool:
     if try_device == "cuda":
         return torch.cuda.is_available()