diff --git a/lerobot/common/optim/schedulers.py b/lerobot/common/optim/schedulers.py
index 7e158394..e2ebb9e3 100644
--- a/lerobot/common/optim/schedulers.py
+++ b/lerobot/common/optim/schedulers.py
@@ -111,6 +111,32 @@ class CosineDecayWithWarmupSchedulerConfig(LRSchedulerConfig):
return LambdaLR(optimizer, lr_lambda, -1)
+@LRSchedulerConfig.register_subclass("constant_with_warmup")
+@dataclass
+class ConstantWithWarmupSchedulerConfig(LRSchedulerConfig):
+ """Used by DexVLA to train Stage2"""
+
+ num_warmup_steps: int
+
+ def build(self, optimizer: Optimizer, num_training_steps: int) -> LambdaLR:
+ 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 constant_schedule(current_step):
+ return 1
+
+ if current_step < self.num_warmup_steps:
+ return linear_warmup_schedule(current_step)
+
+ return constant_schedule(current_step)
+
+ return LambdaLR(optimizer, lr_lambda, -1)
+
+
def save_scheduler_state(scheduler: LRScheduler, save_dir: Path) -> None:
state_dict = scheduler.state_dict()
write_json(state_dict, save_dir / SCHEDULER_STATE)
diff --git a/lerobot/common/policies/__init__.py b/lerobot/common/policies/__init__.py
index b73ba5f4..00e28269 100644
--- a/lerobot/common/policies/__init__.py
+++ b/lerobot/common/policies/__init__.py
@@ -13,6 +13,7 @@
# limitations under the License.
from .act.configuration_act import ACTConfig as ACTConfig
+from .dexvla.configuration_dexvla import DexVLAConfig as DexVLAConfig
from .diffusion.configuration_diffusion import DiffusionConfig as DiffusionConfig
from .pi0.configuration_pi0 import PI0Config as PI0Config
from .tdmpc.configuration_tdmpc import TDMPCConfig as TDMPCConfig
diff --git a/lerobot/common/policies/dexvla/README.md b/lerobot/common/policies/dexvla/README.md
new file mode 100644
index 00000000..cbf94d8b
--- /dev/null
+++ b/lerobot/common/policies/dexvla/README.md
@@ -0,0 +1,140 @@
+
+DexVLA: Vision-Language Model with Plug-In Diffusion Expert for Visuomotor Policy Learning
+
+This policy is Community Contributed. For more information about DexVLA, you can also refer to [this](https://github.com/juruobenruo/DexVLA).
+This is [project website](https://dex-vla.github.io/).
+
+## Dataset
+### Data format
+DexVLA takes RGB images, language instructions and states. For our setting, we use three camera views, namely a top camera and two wrist cameras.
+
+⭐A major difference between DexVLA and other VLAs is: DexVLA takes in raw language, and outputs sub-step reasoning based on current observations.
+So you have to add sub-step reasoning in your data for training.
+
+Specifically, your data should include a key ``reasoning`` which is a list of sub-step reasoning corresponding to each observation.
+For example, if the episode is 10 steps. The length of this list should be 10 as well. And it may looks like:
+~~~python
+reasoning = [
+ "This is step 1.",
+ "This is step 1.",
+ "This is step 2.",
+ "This is step 2.",
+ ...
+ "This is step 4.",
+]
+~~~
+
+Besides, your data should include another key ``action_is_pad`` which is a bool mask indicating whether this action chunk is padded.
+Suppose the size of the action chunk is 5, and the length of the episode is 10. So the action chunk for the last 4 actions must be padded to make sure the length of action chunk is 5.
+And the mask looks like:
+~~~python
+The 6th chunk: [false, false, false, false, true]
+The 7th chunk: [false, false, false, true, true]
+The 8th chunk: [false, false, true, true, true]
+The 9th chunk: [false, true, true, true, true]
+~~~
+
+### Training Data for DexVLA
+The pretraining dataset comprises approximately 100 hours of collected data by ourselves. The dataset mainly including four embodiments which are: moblie Agilex Aloha, single Franka Emika and single UR5e.
+We haven't use any public dataset such as Open-X or DROID.
+
+## 🤗Download Pretrained Weights
+### Download official Qwen2_VL weights
+We construct the VLM backbone by integrating Qwen2-VL-2B, a powerful and efficient model, into our framework.
+The Qwen2-VL 2B serves as the core of our architecture, providing robust capabilities
+for vision-language tasks. We use off-the-shelf Qwen2-VL model proposed
+in [Qwen2-VL](https://arxiv.org/pdf/2409.12191) without any post training on VLM itself. You can download the official weights from this link:
+
+| Model | Link |
+|---------------------|----------------------------------------------------------------|
+| Qwen2-VL (~2B) | [huggingface](https://huggingface.co/Qwen/Qwen2-VL-2B-Instruct) |
+
+**❗❗** After downloading the standard weights, you have to replace the official "config.json"
+with our ["config.json"](https://github.com/juruobenruo/DexVLA/blob/main/docs/config.json) designed for VLA.
+### Download our pretrained ScaleDP-H weights(Stage 1)
+We released our pretrained weights of ScaleDP-H which is trained after Stage1. Now you can download the weights and directly finetuning your data on Stage 2.
+
+| Model | Link |
+|-------------------|----------------------------------------------------------------|
+| ScaleDP-H (~1B) | [huggingface](https://huggingface.co/lesjie/scale_dp_h) |
+| ScaleDP-L (~400M) | [huggingface](https://huggingface.co/lesjie/scale_dp_l) |
+
+**❗❗**After downloading the weights, you have to transform it into ``safetensors`` format, you can simply run this code:
+~~~python
+import torch
+from safetensors.torch import save_file
+path = "/path/to/open_scale_dp_l_backbone.ckpt"
+checkpoint = torch.load(path, map_location=torch.device('cpu'))['nets']['nets']
+
+# Save the weights in safetensors format
+safetensors_path = "/path/to/open_scale_dp_l_backbone.safetensors"
+save_file(checkpoint, safetensors_path)
+print(f"Converted {path} to {safetensors_path}")
+pass
+
+~~~
+
+## 🦾Train
+We have already provided pretrained weights of ScaleDP which is stage 1. Belows are mainly about training process of Stage2 and Stage3.
+
+### Training Stage 2
+~~~shell
+python lerobot/scripts/train.py \
+--policy.type dexvla \
+--policy.qwen2_vl_path /path/to/official/Qwen2-VL-2B-Instruct \
+--policy.pretrain_scaledp_path /path/to/pretrained/scale_dp_h/open_scale_dp_l_backbone.safetensors \
+--policy.policy_head_size 'scaledp_h' \
+--policy.training_stage 2 \
+--dataset.repo_i lerobot/aloha_mobile_chair \
+--policy.using_film true \
+--output_dir /path/to/output \
+--steps 10000 \
+--save_freq 1000 \
+--optimizer_lr 2e-5
+~~~
+
+### Training Stage 3
+Stage3 can be viewed as continual training on specific dexterous tasks like laundry folding which is same as PI0. So stage3 is trained based on stage2.
+~~~shell
+python lerobot/scripts/train.py \
+--policy.type dexvla \
+--policy.qwen2_vl_path /path/to/official/Qwen2-VL-2B-Instruct \
+--.pretrained_path /path/to/pretrained/stage2/weights \
+--policy.policy_head_size 'scaledp_h' \
+--policy.training_stage 3 \
+--dataset.repo_i lerobot/aloha_mobile_chair \
+--batch_size 2 \
+--policy.using_film true \
+--output_dir /path/to/output \
+--steps 10000 \
+--save_freq 1000 \
+--optimizer_lr 2e-5
+~~~
+
+### Training Time
+Original DexVLA is trained on 8 x H100 GPUs. And the training time for each stage is listed as follows:
+
+| Stage | Batch Size(each gpu) | Steps | Time(hour) |
+|--------|----------------------|--------|------------|
+| Stage1 | 32 | 60000 | 30 |
+| Stage2 | 12 | 100000 | 30 |
+| Stage3 | 12 | 60000 | 18 |
+
+
+## Evaluation
+### Evaluation Script
+You can evaluate dexvla by following scripts.
+~~~shell
+python lerobot/scripts/eval.py \
+--policy.type dexvla \
+--policy.pretrained_path /path/to/pretrained/stage2/or/stage3/weights \
+--env.type aloha \
+--env.episode_length 5 \
+--policy.qwen2_vl_path /path/to/official/Qwen2-VL-2B-Instruct \
+--env.task AlohaInsertion-v0 \
+--eval.n_episodes 1 \
+--eval.batch_size 1
+~~~
+
+### Inference Speed
+Tested on a single A6000 GPU, the DexVLA could infer 3.4 action chunks in one second. For each action chunk, if we execute 25 actions, the real control frequency can be 85 (3.4*25)Hz.
diff --git a/lerobot/common/policies/dexvla/configuration_dexvla.py b/lerobot/common/policies/dexvla/configuration_dexvla.py
new file mode 100644
index 00000000..96a3944b
--- /dev/null
+++ b/lerobot/common/policies/dexvla/configuration_dexvla.py
@@ -0,0 +1,179 @@
+#!/usr/bin/env python
+
+# Copyright 2025 DexVLA Team 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.
+
+"""Qwen2VL model configuration"""
+
+from dataclasses import dataclass, field
+from typing import Tuple
+
+from transformers import AutoConfig
+from transformers.utils import logging
+
+from lerobot.common.optim.optimizers import AdamWConfig
+from lerobot.common.optim.schedulers import (
+ ConstantWithWarmupSchedulerConfig,
+ CosineDecayWithWarmupSchedulerConfig,
+)
+from lerobot.configs.policies import PreTrainedConfig
+from lerobot.configs.types import NormalizationMode
+
+from .policy_heads import register_policy_heads
+from .qwe2_vla import register_qwen2_vla
+
+logger = logging.get_logger(__name__)
+register_policy_heads()
+register_qwen2_vla()
+
+
+@PreTrainedConfig.register_subclass("dexvla")
+@dataclass
+class DexVLAConfig(PreTrainedConfig):
+ # For loading policy head
+ policy_head_type: str = "scale_dp_policy"
+ policy_head_size: str = "scaledp_l"
+ action_dim: int = 14
+ state_dim: int = 14
+ chunk_size: int = 50
+ n_action_steps: int = 50
+ n_obs_steps: int = 1
+
+ device: str = "cuda"
+
+ hidden_size: int = 1536
+ qwen2_vl_path: str = (
+ None # '/media/rl/HDD/data/weights/Qwen2-VL-2B-Instruct', official weights of qwen2vl
+ )
+
+ pretrained_path: str = None # for loading pretrained weights of whole dexvla, usually for training stage3
+ pretrained_scaledp_path: str = None # for loading pretrained weights of ScaleDP(Stage1)
+
+ training_stage: int = 2 # specific training stage, [2, 3]
+ using_film: bool = True
+ llm_loss_weight: float = 1.0
+ with_llm_head: bool = True
+ using_reasoning: bool = True
+ resize_size: tuple = (240, 320)
+ # Training presets
+ optimizer_lr: float = 2e-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 = 2_000
+ scheduler_decay_steps: int = 30_000
+ scheduler_decay_lr: float = 2.5e-6
+
+ normalization_mapping: dict[str, NormalizationMode] = field(
+ default_factory=lambda: {
+ # "VISUAL": NormalizationMode.MEAN_STD,
+ "STATE": NormalizationMode.MEAN_STD,
+ "ACTION": NormalizationMode.MIN_MAX,
+ }
+ )
+
+ def __post_init__(self):
+ 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.using_reasoning:
+ assert self.using_film, "using_reasoning requires `using_film=True`"
+ assert self.with_llm_head, "using_reasoning requires `with_llm_head=True`"
+ print("You have set using_reasoning=True, please make sure your data has key 'reasoning'.")
+ else:
+ print(
+ "Warning:DexVLA recommends to use reasoning data which can better handle long-horizon and dexterous tasks. You can set 'using_reaasoning=True'."
+ )
+
+ if self.qwen2_vl_path is None:
+ raise ValueError(
+ "DexVLA is built on official qwen2_vl-2B. You have to download the official weights of qwen2_vl-2B first and set 'qwen2_vl_path'."
+ )
+
+ if self.policy_head_type == "scale_dp_policy":
+ self.policy_head_config = AutoConfig.for_model(
+ model_type=self.policy_head_type,
+ model_size=self.policy_head_size,
+ cond_dim=self.hidden_size,
+ action_dim=self.action_dim,
+ prediction_horizon=self.chunk_size,
+ state_dim=self.state_dim,
+ )
+ elif self.policy_head_type == "unet_diffusion":
+ self.policy_head_config = AutoConfig.for_model(
+ model_type=self.policy_head_type,
+ global_cond_dim=self.hidden_size,
+ action_dim=self.action_dim,
+ state_dim=self.state_dim,
+ )
+ else:
+ raise ValueError(f"Policy head type {self.policy_head_type} not supported")
+
+ if self.training_stage not in [2, 3]:
+ raise ValueError(f"Training stage must be 2 or 3. Got {self.training_stage}.")
+
+ self.qwen2_vla_config = AutoConfig.from_pretrained(self.qwen2_vl_path)
+
+ 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):
+ if self.training_stage == 3:
+ 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,
+ )
+ else:
+ return ConstantWithWarmupSchedulerConfig(
+ num_warmup_steps=self.scheduler_warmup_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
diff --git a/lerobot/common/policies/dexvla/fusion_modules.py b/lerobot/common/policies/dexvla/fusion_modules.py
new file mode 100644
index 00000000..39bbc57f
--- /dev/null
+++ b/lerobot/common/policies/dexvla/fusion_modules.py
@@ -0,0 +1,58 @@
+#!/usr/bin/env python
+
+# Copyright 2025 DexVLA Team 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.
+
+import torch.nn as nn
+
+
+class ActionProjector(nn.Module):
+ def __init__(self, in_dim, out_dim=1024):
+ super().__init__()
+ self.global_1d_pool = nn.AdaptiveAvgPool1d(1)
+ self.mlps = nn.ModuleList(
+ [
+ # nn.LayerNorm(in_dim),
+ nn.Linear(in_dim, in_dim),
+ nn.GELU(),
+ nn.Linear(in_dim, out_dim),
+ nn.Dropout(0.0),
+ ]
+ )
+
+ def forward(self, x):
+ x = self.global_1d_pool(x.permute(1, 0)).permute(1, 0)
+ for mlp in self.mlps:
+ x = mlp(x)
+ return x
+
+
+class FiLM(nn.Module):
+ def __init__(self, feature_dim, condition_dim):
+ super().__init__()
+ self.scale_fc = nn.Linear(condition_dim, feature_dim)
+ self.shift_fc = nn.Linear(condition_dim, feature_dim)
+
+ nn.init.zeros_(self.scale_fc.weight)
+ nn.init.zeros_(self.scale_fc.bias)
+ nn.init.zeros_(self.shift_fc.weight)
+ nn.init.zeros_(self.shift_fc.bias)
+
+ def forward(self, x, condition):
+ # calculate scale and shift
+ scale = self.scale_fc(condition)
+ shift = self.shift_fc(condition)
+
+ # film
+ return x * (1 + scale) + shift
diff --git a/lerobot/common/policies/dexvla/modeling_dexvla.py b/lerobot/common/policies/dexvla/modeling_dexvla.py
new file mode 100644
index 00000000..e1133df8
--- /dev/null
+++ b/lerobot/common/policies/dexvla/modeling_dexvla.py
@@ -0,0 +1,313 @@
+#!/usr/bin/env python
+
+# Copyright 2025 DexVLA Team 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.
+
+from collections import deque
+
+import torch
+import torchvision.transforms as transforms
+from safetensors.torch import load_file
+from torch import Tensor
+from transformers import AutoModelForCausalLM, AutoProcessor, AutoTokenizer
+
+from lerobot.common.policies.dexvla.configuration_dexvla import DexVLAConfig
+from lerobot.common.policies.dexvla.robot_data_processor import Qwen2VLAProcess
+from lerobot.common.policies.normalize import Normalize, Unnormalize
+from lerobot.common.policies.pretrained import PreTrainedPolicy
+
+
+class DexVLAPolicy(PreTrainedPolicy):
+ """Wrapper class around Qwen2VLForConditionalGenerationForVLA model to train and run inference within LeRobot."""
+
+ config_class = DexVLAConfig
+ name = "dexvla"
+
+ def __init__(
+ self,
+ config: DexVLAConfig,
+ 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
+ )
+
+ for k in ["using_film", "llm_loss_weight", "with_llm_head", "policy_head_config"]:
+ setattr(config.qwen2_vla_config, k, config.__dict__[k])
+
+ # if self.config.training_stage == 2:
+ # self.model = Qwen2VLForConditionalGenerationForVLA(config.qwen2_vla_config).to(torch.bfloat16)
+ model_base = self.config.qwen2_vl_path
+ self.model = AutoModelForCausalLM.from_pretrained(
+ model_base,
+ config=config.qwen2_vla_config,
+ trust_remote_code=True,
+ _fast_init=False,
+ # attn_implementation="flash_attention_2",
+ ).to(device="cuda", dtype=torch.bfloat16)
+
+ if self.config.pretrained_scaledp_path is not None:
+ print(
+ "\n>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>>Loading pretrained ScaleDP weights...<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<<"
+ )
+ pretrain_scaledp_weights = load_file(self.config.pretrained_scaledp_path)
+
+ keys_to_del_dit = []
+ pretrain_scaledp_weights = {
+ k[7:] if k.startswith("policy.") else k: v for k, v in pretrain_scaledp_weights.items()
+ }
+ for k in pretrain_scaledp_weights:
+ if "noise_pred" not in k: # del weights of vision backbones
+ keys_to_del_dit.append(k)
+ if "cond_obs_emb" in k:
+ keys_to_del_dit.append(k)
+ for k in keys_to_del_dit:
+ del pretrain_scaledp_weights[k]
+ pretrain_scaledp_weights = {
+ k[15:] if k.startswith("noise_pred_net.") else k: v
+ for k, v in pretrain_scaledp_weights.items()
+ }
+
+ self.model.policy_head.load_state_dict(pretrain_scaledp_weights, strict=False)
+
+ self.model.requires_grad_(False)
+ self.model.policy_head.requires_grad_(True)
+ self.qwen2_vl_processor = AutoProcessor.from_pretrained(config.qwen2_vl_path)
+ self.tokenizer = AutoTokenizer.from_pretrained(config.qwen2_vl_path)
+ self.vla_processor = Qwen2VLAProcess(
+ tokenizer=self.tokenizer, multimodal_processor=self.qwen2_vl_processor
+ ) # process the input data into VLM format
+
+ self.resize_size = self.config.resize_size
+ ratio = 0.95
+ self.transformations = [
+ transforms.Resize(size=self.resize_size, antialias=True),
+ transforms.RandomCrop(size=[int(self.resize_size[0] * ratio), int(self.resize_size[1] * ratio)]),
+ transforms.Resize(self.resize_size, antialias=True),
+ transforms.RandomRotation(degrees=[-5.0, 5.0], expand=False),
+ transforms.ColorJitter(brightness=0.3, contrast=0.4, saturation=0.5), # , hue=0.08)
+ ]
+
+ self.reset()
+
+ def process_batch(self, batch: dict[str, Tensor]) -> dict[str, Tensor]:
+ """Applying DexVLA preprocessing to original data. Including resizing images. Scaling the range of actions, states."""
+ batch = self.normalize_inputs(batch)
+ batch = self.normalize_targets(batch)
+ present_img_keys = [key for key in self.config.image_features if key in batch]
+ task_descs = batch["task"]
+ try:
+ reasonings = batch["reasoning"]
+ except KeyError:
+ reasonings = ["None."] * len(task_descs)
+
+ pass
+ is_pad = batch["action_is_pad"]
+ all_cam_images = []
+ for k in present_img_keys:
+ all_cam_images.append(batch[k])
+
+ # construct observations, and scale 0-1 to 0-255
+ image_data = torch.stack(all_cam_images) * 255
+ image_data = image_data.to(dtype=torch.uint8)
+ # construct observations
+ qpos_data = batch["observation.state"].float()
+ action_data = batch["action"].float()
+
+ orig_shape = image_data.shape
+ image_data = image_data.view(-1, *orig_shape[2:])
+
+ for transform in self.transformations:
+ image_data = transform(image_data)
+
+ image_data = image_data.view(*orig_shape[:3], *self.resize_size)
+
+ vl_data = {"images": image_data, "raw_langs": task_descs, "reasonings": reasonings}
+ # processing vl_data into qwen2_vl format
+ vla_inputs = self.vla_processor.forward(vl_data, use_reasoning=self.config.using_reasoning)
+ vla_inputs["states"] = qpos_data
+ vla_inputs["is_pad"] = is_pad
+ vla_inputs["actions"] = action_data
+ return vla_inputs
+
+ def forward(self, batch: dict[str, Tensor]) -> tuple[Tensor, dict[str, Tensor]]:
+ processed_batch = self.process_batch(batch)
+
+ ret = self.model.forward(**processed_batch)
+ loss_dict = ret["loss"]
+ loss = loss_dict["loss"].mean()
+ return loss, loss_dict
+
+ def dexvla_predict_action(
+ self,
+ input_ids: torch.LongTensor = None,
+ actions=None,
+ states=None,
+ is_pad=None,
+ tokenizer=None,
+ is_eval=True,
+ pixel_values=None,
+ attention_mask=None,
+ image_grid_spatiotemporal=None,
+ ):
+ input_ids = input_ids.to("cuda")
+ with torch.inference_mode():
+ outputs = self.model.generate(
+ input_ids,
+ pixel_values=pixel_values,
+ attention_mask=attention_mask,
+ image_grid_spatiotemporal=image_grid_spatiotemporal,
+ is_eval=is_eval,
+ num_beams=1,
+ do_sample=False,
+ temperature=0.2,
+ max_new_tokens=256,
+ eos_token_id=tokenizer.eos_token_id, # End of sequence token
+ pad_token_id=tokenizer.eos_token_id, # Pad token
+ use_cache=True,
+ output_hidden_states=True,
+ return_dict_in_generate=True,
+ )
+
+ output_ids = outputs.sequences
+ # last_hidden_states = outputs.hidden_states[-2][-1]
+ input_token_len = input_ids.shape[1]
+ n_diff_input_output = (input_ids != output_ids[:, :input_token_len]).sum().item()
+ if n_diff_input_output > 0:
+ print(f"[Warning] {n_diff_input_output} output_ids are not the same as the input_ids")
+ outputs_text = tokenizer.batch_decode(output_ids[:, input_token_len:], skip_special_tokens=False)[0]
+
+ outputs_text = outputs_text.strip()
+ last_hidden_states = [each[-1] for each in outputs.hidden_states] # all hidden states
+ all_hidden_states = torch.cat(last_hidden_states, dim=1)
+
+ action_hidden_states = None
+ labels_input = torch.ones((1, input_token_len)) * -100
+ labels_output = torch.ones((1, output_ids.shape[1] - input_token_len))
+ labels = torch.cat([labels_input, labels_output], dim=1)
+
+ if self.model.using_film:
+ action_hidden_states = self.model.film_forward(
+ labels=labels,
+ input_ids=output_ids,
+ hidden_states=torch.cat(last_hidden_states, dim=1),
+ )
+
+ action = self.model.policy_head(
+ actions, action_hidden_states, states.to(all_hidden_states.dtype), is_pad
+ )
+ return action, outputs_text
+
+ def tinyvla_predict_action(
+ self,
+ input_ids: torch.LongTensor = None,
+ actions=None,
+ states=None,
+ is_pad=None,
+ is_eval=True,
+ pixel_values=None,
+ attention_mask=None,
+ image_grid_spatiotemporal=None,
+ ):
+ input_ids = input_ids.to("cuda")
+ with torch.inference_mode():
+ all_hidden_states = self.model.forward(
+ input_ids,
+ pixel_values=pixel_values,
+ attention_mask=attention_mask,
+ image_grid_spatiotemporal=image_grid_spatiotemporal,
+ is_eval=is_eval,
+ tinyvla=True,
+ )
+
+ all_hidden_states = torch.mean(all_hidden_states, dim=1).unsqueeze(1)
+
+ action = self.model.policy_head(
+ actions, all_hidden_states, states.to(all_hidden_states.dtype), is_pad
+ )
+ return action, "tinyvla generates no reasoning"
+
+ 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()
+ batch = self.normalize_inputs(batch)
+
+ if len(self._action_queue) == 0:
+ present_img_keys = [key for key in self.config.image_features if key in batch]
+ try:
+ task_descs = batch["task"]
+ except KeyError:
+ task_descs = " "
+ print("No task descriptions found for this task")
+
+ all_cam_images = []
+ for k in present_img_keys:
+ all_cam_images.append(batch[k])
+
+ # construct observations, and scale 0-1 to 0-255
+ image_data = torch.stack(all_cam_images) * 255
+ image_data = image_data.to(dtype=torch.uint8)
+ # construct observations
+ qpos_data = batch["observation.state"].float()
+
+ image_data = image_data.squeeze(0)
+
+ for transform in self.transformations:
+ image_data = transform(image_data)
+
+ # processing vl_data into qwen2_vl format
+ vla_inputs = self.vla_processor.single_forward_process(
+ images=image_data, raw_lang=task_descs, reasoning=None, eval=True
+ )
+ vla_inputs["states"] = qpos_data
+
+ if self.config.using_film and self.config.with_llm_head: # dexvla
+ all_actions, outputs = self.dexvla_predict_action(
+ **vla_inputs, is_eval=True, tokenizer=self.tokenizer
+ )
+ else: # tinyvla
+ all_actions, outputs = self.tinyvla_predict_action(**vla_inputs, is_eval=True)
+
+ actions = self.unnormalize_outputs({"action": all_actions})["action"]
+ self._action_queue.extend(actions.transpose(0, 1))
+
+ return self._action_queue.popleft()
diff --git a/lerobot/common/policies/dexvla/policy_heads/__init__.py b/lerobot/common/policies/dexvla/policy_heads/__init__.py
new file mode 100644
index 00000000..f3b6a169
--- /dev/null
+++ b/lerobot/common/policies/dexvla/policy_heads/__init__.py
@@ -0,0 +1,29 @@
+#!/usr/bin/env python
+
+# Copyright 2025 DexVLA Team 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.
+
+from transformers import AutoConfig, AutoModel
+
+from .configuration_scaledp import ScaleDPPolicyConfig
+from .configuration_unet_diffusion import UnetDiffusionPolicyConfig
+from .modeling_scaledp import ScaleDP
+from .modeling_unet_diffusion import ConditionalUnet1D
+
+
+def register_policy_heads():
+ AutoConfig.register("scale_dp_policy", ScaleDPPolicyConfig)
+ AutoConfig.register("unet_diffusion_policy", UnetDiffusionPolicyConfig)
+ AutoModel.register(ScaleDPPolicyConfig, ScaleDP)
+ AutoModel.register(UnetDiffusionPolicyConfig, ConditionalUnet1D)
diff --git a/lerobot/common/policies/dexvla/policy_heads/configuration_scaledp.py b/lerobot/common/policies/dexvla/policy_heads/configuration_scaledp.py
new file mode 100644
index 00000000..e2d71cea
--- /dev/null
+++ b/lerobot/common/policies/dexvla/policy_heads/configuration_scaledp.py
@@ -0,0 +1,123 @@
+#!/usr/bin/env python
+
+# Copyright 2025 DexVLA Team 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.
+
+
+import os
+from typing import Union
+
+from transformers import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+MODEL_STRUCTURE = {
+ "scaledp_h": {
+ "depth": 32,
+ "n_emb": 1280,
+ "num_heads": 16,
+ },
+ "scaledp_l": {
+ "depth": 24,
+ "n_emb": 1024,
+ "num_heads": 16,
+ }, # 400M
+}
+
+
+class ScaleDPPolicyConfig(PretrainedConfig):
+ """
+ Configuration for ScaleDP policy head
+ """
+
+ model_type = "scale_dp_policy"
+
+ def __init__(
+ self,
+ eval: bool = False,
+ action_dim: int = 14, # action dim
+ # output_dim: int = 14, # action dim
+ cond_dim: int = 1536, # the input dim of the condition
+ state_dim: int = 14, # the input dim of the state
+ prediction_horizon: int = 16, # horizon
+ n_obs_steps: int = 2, # number of observation steps
+ depth: int = 28, # number of DiT blocks
+ n_emb: int = 256, # embedding size
+ num_heads: int = 16,
+ mlp_ratio: int = 4.0,
+ time_as_cond: bool = True,
+ obs_as_cond: bool = True,
+ learn_sigma: bool = False,
+ model_size: str = "none",
+ num_inference_timesteps: int = 10,
+ noise_samples: int = 1,
+ num_train_timesteps: int = 100,
+ **kwargs,
+ ):
+ if model_size != "none":
+ depth = MODEL_STRUCTURE[model_size]["depth"]
+ n_emb = MODEL_STRUCTURE[model_size]["n_emb"]
+ num_heads = MODEL_STRUCTURE[model_size]["num_heads"]
+ else:
+ # raise ValueError("model_size show not be 'none'")
+ pass
+ # print("model_size should not be 'none'")
+ self.eval = eval
+
+ self.input_dim = action_dim
+ self.output_dim = action_dim
+ self.prediction_horizon = prediction_horizon
+
+ self.cond_dim = cond_dim
+ self.state_dim = state_dim
+
+ self.n_obs_steps = n_obs_steps
+ self.depth = depth
+ self.n_emb = n_emb
+ self.num_heads = num_heads
+ self.mlp_ratio = mlp_ratio
+ self.time_as_cond = time_as_cond
+ self.obs_as_cond = obs_as_cond
+ self.learn_sigma = learn_sigma
+
+ self.num_inference_timesteps = num_inference_timesteps
+ self.num_queries = prediction_horizon
+ self.noise_samples = noise_samples
+ self.num_train_timesteps = num_train_timesteps
+ super().__init__(**kwargs)
+
+ @classmethod
+ def from_pretrained(
+ cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs
+ ) -> "PretrainedConfig":
+ cls._set_token_in_kwargs(kwargs)
+
+ config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+ # get the vision config dict if we are loading from CLIPConfig
+ if config_dict.get("model_type") == "llava_pythia":
+ config_dict = config_dict["action_head"]
+
+ if (
+ "model_type" in config_dict
+ and hasattr(cls, "model_type")
+ and config_dict["model_type"] != cls.model_type
+ ):
+ logger.warning(
+ f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+ f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+ )
+
+ return cls.from_dict(config_dict, **kwargs)
diff --git a/lerobot/common/policies/dexvla/policy_heads/configuration_unet_diffusion.py b/lerobot/common/policies/dexvla/policy_heads/configuration_unet_diffusion.py
new file mode 100644
index 00000000..b7eb046e
--- /dev/null
+++ b/lerobot/common/policies/dexvla/policy_heads/configuration_unet_diffusion.py
@@ -0,0 +1,86 @@
+#!/usr/bin/env python
+
+# Copyright 2025 DexVLA Team 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.
+
+import os
+from typing import Union
+
+from transformers import PretrainedConfig
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class UnetDiffusionPolicyConfig(PretrainedConfig):
+ """
+ Configuration for dit diffusion policy head
+ """
+
+ model_type = "unet_diffusion_policy"
+
+ def __init__(
+ self,
+ action_dim=10,
+ global_cond_dim=2048,
+ diffusion_step_embed_dim=256,
+ down_dims=None,
+ kernel_size=5,
+ n_groups=8,
+ state_dim=7,
+ prediction_horizon=16,
+ noise_samples=1,
+ num_inference_timesteps=10,
+ num_train_timesteps=100,
+ **kwargs,
+ ):
+ if down_dims is None:
+ down_dims = [256, 512, 1024]
+ self.input_dim = action_dim
+ self.noise_samples = noise_samples
+ self.prediction_horizon = prediction_horizon
+ self.num_inference_timesteps = num_inference_timesteps
+ self.global_cond_dim = global_cond_dim
+ self.diffusion_step_embed_dim = diffusion_step_embed_dim
+ self.down_dims = down_dims
+ self.kernel_size = kernel_size
+ self.n_groups = n_groups
+ self.state_dim = state_dim
+ self.num_train_timesteps = num_train_timesteps
+
+ super().__init__(**kwargs)
+
+ @classmethod
+ def from_pretrained(
+ cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs
+ ) -> "PretrainedConfig":
+ cls._set_token_in_kwargs(kwargs)
+
+ config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+ # get the vision config dict if we are loading from CLIPConfig
+ if config_dict.get("model_type") == "llava_pythia":
+ config_dict = config_dict["action_head"]
+
+ if (
+ "model_type" in config_dict
+ and hasattr(cls, "model_type")
+ and config_dict["model_type"] != cls.model_type
+ ):
+ logger.warning(
+ f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+ f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+ )
+
+ return cls.from_dict(config_dict, **kwargs)
diff --git a/lerobot/common/policies/dexvla/policy_heads/modeling_scaledp.py b/lerobot/common/policies/dexvla/policy_heads/modeling_scaledp.py
new file mode 100644
index 00000000..62ff0587
--- /dev/null
+++ b/lerobot/common/policies/dexvla/policy_heads/modeling_scaledp.py
@@ -0,0 +1,561 @@
+#!/usr/bin/env python
+
+# Copyright 2025 DexVLA Team 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.
+
+import logging
+import math
+from typing import Tuple
+
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as func
+import torch.utils.checkpoint
+from timm.models.vision_transformer import Mlp, use_fused_attn
+from torch.jit import Final
+from transformers.modeling_utils import PreTrainedModel
+
+from .configuration_scaledp import ScaleDPPolicyConfig
+
+_logger = logging.getLogger(__name__)
+
+
+class Attention(nn.Module):
+ fused_attn: Final[bool]
+
+ def __init__(
+ self,
+ dim: int,
+ num_heads: int = 8,
+ qkv_bias: bool = False,
+ qk_norm: bool = False,
+ attn_drop: float = 0.0,
+ proj_drop: float = 0.0,
+ norm_layer: nn.Module = nn.LayerNorm,
+ ) -> None:
+ super().__init__()
+ assert dim % num_heads == 0, "dim should be divisible by num_heads"
+ self.num_heads = num_heads
+ self.head_dim = dim // num_heads
+ self.scale = self.head_dim**-0.5
+ self.fused_attn = use_fused_attn()
+
+ self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+ self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
+ self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
+ self.attn_drop = nn.Dropout(attn_drop)
+ self.proj = nn.Linear(dim, dim)
+ self.proj_drop = nn.Dropout(proj_drop)
+
+ def forward(self, x: torch.Tensor, attn_mask=None) -> torch.Tensor:
+ b, n, c = x.shape
+ qkv = self.qkv(x).reshape(b, n, 3, self.num_heads, self.head_dim).permute(2, 0, 3, 1, 4)
+ q, k, v = qkv.unbind(0)
+ q, k = self.q_norm(q), self.k_norm(k)
+
+ if self.fused_attn:
+ x = func.scaled_dot_product_attention(
+ q,
+ k,
+ v,
+ attn_mask=attn_mask,
+ dropout_p=self.attn_drop.p if self.training else 0.0,
+ )
+ else:
+ q = q * self.scale
+ # attn = q @ k.transpose(-2, -1)
+ # if attn_mask is not None:
+ # attn += attn_mask
+ # attn = attn.softmax(dim=-1)
+ # attn = self.attn_drop(attn)
+ # x = attn @ v
+ attn_scores = torch.matmul(q, k.transpose(-2, -1))
+
+ # Add attention mask if provided
+ if attn_mask is not None:
+ attn_scores += attn_mask
+
+ # Apply softmax to get attention weights (softmax is applied along the last dimension)
+ attn_weights = func.softmax(attn_scores, dim=-1)
+
+ # Dropout on attention weights (if dropout is used)
+ attn_weights = self.attn_drop(attn_weights)
+
+ # Apply attention weights to value tensor (V)
+ x = torch.matmul(attn_weights, v)
+
+ x = x.transpose(1, 2).reshape(b, n, c)
+ x = self.proj(x)
+ x = self.proj_drop(x)
+ return x
+
+
+logger = logging.getLogger(__name__)
+
+
+def modulate(x, shift, scale):
+ return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)
+
+
+#################################################################################
+# Embedding Layers for Timesteps and Class Labels #
+#################################################################################
+
+
+class TimestepEmbedder(nn.Module):
+ """
+ Embeds scalar timesteps into vector representations.
+ """
+
+ def __init__(self, hidden_size, frequency_embedding_size=256):
+ super().__init__()
+ self.mlp = nn.Sequential(
+ nn.Linear(frequency_embedding_size, hidden_size, bias=True),
+ nn.SiLU(),
+ nn.Linear(hidden_size, hidden_size, bias=True),
+ )
+ self.frequency_embedding_size = frequency_embedding_size
+
+ @staticmethod
+ def timestep_embedding(t, dim, max_period=10000):
+ """
+ Create sinusoidal timestep embeddings.
+ :param t: a 1-D Tensor of N indices, one per batch element.
+ These may be fractional.
+ :param dim: the dimension of the output.
+ :param max_period: controls the minimum frequency of the embeddings.
+ :return: an (N, D) Tensor of positional embeddings.
+ """
+ # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
+ half = dim // 2
+ freqs = torch.exp(
+ -math.log(max_period) * torch.arange(start=0, end=half, dtype=torch.bfloat16) / half
+ ).to(device=t.device)
+ args = t[:, None].float() * freqs[None]
+ embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+ if dim % 2:
+ embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+ return embedding.to(dtype=torch.bfloat16)
+
+ def forward(self, t):
+ t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
+ t_emb = self.mlp(t_freq)
+ return t_emb
+
+
+#################################################################################
+# Core ScaleDP Model #
+#################################################################################
+
+
+class ScaleDPBlock(nn.Module):
+ """
+ A ScaleDP block with adaptive layer norm zero (adaLN-Zero) conScaleDPioning.
+ """
+
+ def __init__(self, hidden_size, num_heads, mlp_ratio=4.0, **block_kwargs):
+ super().__init__()
+ self.norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+ self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=True, **block_kwargs)
+ self.norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+ mlp_hidden_dim = int(hidden_size * mlp_ratio)
+
+ def approx_gelu():
+ return nn.GELU(approximate="tanh")
+
+ self.mlp = Mlp(in_features=hidden_size, hidden_features=mlp_hidden_dim, act_layer=approx_gelu, drop=0)
+ self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 6 * hidden_size, bias=True))
+
+ def forward(self, x, c, attn_mask=None):
+ shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(c).chunk(
+ 6, dim=1
+ )
+ x = x + gate_msa.unsqueeze(1) * self.attn(
+ modulate(self.norm1(x), shift_msa, scale_msa), attn_mask=attn_mask
+ ) # norm, scale&shift, attn, scale,
+ x = x + gate_mlp.unsqueeze(1) * self.mlp(modulate(self.norm2(x), shift_mlp, scale_mlp))
+ return x
+
+
+class FinalLayer(nn.Module):
+ """
+ The final layer of ScaleDP.
+ """
+
+ def __init__(self, hidden_size, output_dim):
+ super().__init__()
+ self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
+ self.linear = nn.Linear(hidden_size, output_dim, bias=True)
+ self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=True))
+
+ def forward(self, x, c):
+ shift, scale = self.adaLN_modulation(c).chunk(2, dim=1)
+ x = modulate(self.norm_final(x), shift, scale)
+ x = self.linear(x)
+ return x
+
+
+class ScaleDP(PreTrainedModel):
+ """
+ Diffusion models with a Transformer backbone.
+ """
+
+ config_class = ScaleDPPolicyConfig
+
+ def __init__(
+ self,
+ config: ScaleDPPolicyConfig,
+ ):
+ super().__init__(config)
+ # compute number of tokens for main trunk and conScaleDPion encoder
+ if config.n_obs_steps is None:
+ config.n_obs_steps = config.prediction_horizon
+ t = config.prediction_horizon
+ t_cond = 1
+ if not config.time_as_cond:
+ t += 1
+ t_cond -= 1
+ obs_as_cond = config.cond_dim > 0
+ if obs_as_cond:
+ assert config.time_as_cond
+ t_cond += config.n_obs_steps
+
+ # self.combine = nn.Linear(cond_dim+state_dim, cond_dim)
+ self.combine = nn.Sequential(
+ nn.Linear(config.cond_dim + config.state_dim, 1024),
+ nn.ReLU(),
+ nn.Linear(1024, 1024),
+ nn.ReLU(),
+ nn.Linear(1024, config.cond_dim),
+ )
+ self.learn_sigma = config.learn_sigma
+ self.input_dim = config.input_dim
+ self.output_dim = config.output_dim * 2 if config.learn_sigma else config.output_dim
+ self.num_heads = config.num_heads
+
+ self.x_embedder = nn.Linear(config.input_dim, config.n_emb)
+ self.t_embedder = TimestepEmbedder(config.n_emb)
+ self.cond_obs_emb = None
+ if obs_as_cond:
+ self.cond_obs_emb = nn.Linear(config.cond_dim, config.n_emb)
+
+ # Will use fixed sin-cos embedding:
+ self.pos_embed = nn.Parameter(torch.zeros(1, config.prediction_horizon, config.n_emb))
+
+ self.blocks = nn.ModuleList(
+ [
+ ScaleDPBlock(config.n_emb, config.num_heads, mlp_ratio=config.mlp_ratio)
+ for _ in range(config.depth)
+ ]
+ )
+ self.final_layer = FinalLayer(config.n_emb, output_dim=config.output_dim)
+ # self.initialize_weights()
+ # constants
+ self.t = t
+ self.t_cond = t_cond
+ self.prediction_horizon = config.prediction_horizon
+ self.time_as_cond = config.time_as_cond
+ self.action_dim = config.output_dim
+ self.obs_as_cond = obs_as_cond
+ logger.info("number of parameters in ScaleDP: %e", sum(p.numel() for p in self.parameters()))
+
+ from diffusers.schedulers.scheduling_ddim import DDIMScheduler
+
+ self.num_inference_timesteps = config.num_inference_timesteps
+ # self.proj_to_action = nn.Identity()
+ self.noise_scheduler = DDIMScheduler(
+ num_train_timesteps=config.num_train_timesteps, # 100
+ beta_schedule="squaredcos_cap_v2",
+ clip_sample=True,
+ set_alpha_to_one=True,
+ steps_offset=0,
+ prediction_type="epsilon",
+ )
+ self.num_queries = config.num_queries # 16
+ self.noise_samples = config.noise_samples # 1
+ # self.num_inference_timesteps = config.num_inference_timesteps # 100
+
+ def initialize_weights(self):
+ # Initialize transformer layers:
+ def _basic_init(module):
+ if isinstance(module, nn.Linear):
+ torch.nn.init.xavier_uniform_(module.weight)
+ if module.bias is not None:
+ nn.init.constant_(module.bias, 0)
+
+ self.apply(_basic_init)
+
+ nn.init.normal_(self.pos_embed, mean=0.0, std=0.02)
+
+ # Initialize patch_embed like nn.Linear (instead of nn.Conv2d):
+ w = self.x_embedder.weight.data
+ nn.init.xavier_uniform_(w.view([w.shape[0], -1]))
+ nn.init.constant_(self.x_embedder.bias, 0)
+
+ # Initialize label embedding table:
+ nn.init.normal_(self.cond_obs_emb.weight, mean=0.0, std=0.02)
+ nn.init.constant_(self.cond_obs_emb.bias, 0)
+
+ # Initialize timestep embedding MLP:
+ nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
+ nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
+
+ # Zero-out adaLN modulation layers in ScaleDP blocks:
+ for block in self.blocks:
+ nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
+ nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
+
+ # Zero-out output layers:
+ nn.init.constant_(self.final_layer.adaLN_modulation[-1].weight, 0)
+ nn.init.constant_(self.final_layer.adaLN_modulation[-1].bias, 0)
+ nn.init.constant_(self.final_layer.linear.weight, 0)
+ nn.init.constant_(self.final_layer.linear.bias, 0)
+
+ def get_optim_groups(self, weight_decay: float = 1e-3):
+ """
+ This long function is unfortunately doing something very simple and is being very defensive:
+ We are separating out all parameters of the models into two buckets: those that will experience
+ weight decay for regularization and those that won't (biases, and layernorm/embedding weights).
+ We are then returning the PyTorch optimizer object.
+ """
+
+ # separate out all parameters to those that will and won't experience regularizing weight decay
+ decay = set()
+ no_decay = set()
+ whitelist_weight_modules = (torch.nn.Linear, Attention)
+ blacklist_weight_modules = (torch.nn.LayerNorm, torch.nn.Embedding)
+ for mn, m in self.named_modules():
+ for pn, _p in m.named_parameters():
+ fpn = "{}.{}".format(mn, pn) if mn else pn # full param name
+
+ if pn.endswith("bias"):
+ # all biases will not be decayed
+ no_decay.add(fpn)
+ elif pn.startswith("bias"):
+ # MultiheadAttention bias starts with "bias"
+ no_decay.add(fpn)
+ elif pn.endswith("weight") and isinstance(m, whitelist_weight_modules):
+ # weights of whitelist modules will be weight decayed
+ decay.add(fpn)
+ elif pn.endswith("weight") and isinstance(m, blacklist_weight_modules):
+ # weights of blacklist modules will NOT be weight decayed
+ no_decay.add(fpn)
+
+ # validate that we considered every parameter
+ param_dict = dict(self.named_parameters())
+ inter_params = decay & no_decay
+ union_params = decay | no_decay
+ assert len(inter_params) == 0, "parameters {} made it into both decay/no_decay sets!".format(
+ str(inter_params)
+ )
+ assert len(param_dict.keys() - union_params) == 0, (
+ "parameters {} were not separated into either decay/no_decay set!".format(
+ str(param_dict.keys() - union_params),
+ )
+ )
+
+ # create the pytorch optimizer object
+ optim_groups = [
+ {
+ "params": [param_dict[pn] for pn in sorted(decay)],
+ "weight_decay": weight_decay,
+ },
+ {
+ "params": [param_dict[pn] for pn in sorted(no_decay)],
+ "weight_decay": 0.0,
+ },
+ ]
+ return optim_groups
+
+ def configure_optimizers(
+ self,
+ learning_rate: float = 1e-4,
+ weight_decay: float = 1e-3,
+ betas: Tuple[float, float] = (0.9, 0.95),
+ ):
+ optim_groups = self.get_optim_groups(weight_decay=weight_decay)
+ optimizer = torch.optim.AdamW(optim_groups, lr=learning_rate, betas=betas)
+ return optimizer
+
+ def forward(self, actions, hidden_states, states, is_pad):
+ """
+ Forward pass for the diffusion head.
+ :param actions: target actions, shape [b, Ta, D] D:10 = 3+6+1
+ :param hidden_states: hidden states from the llava_pythia, as the conScaleDPion for the diffusion, shape [b,Tokens, D] 8 1200 1024
+ :param states: robot states, shape [b, D]
+ :return: loss
+ """
+ if actions is not None: # training time
+ b = actions.size(0)
+ actions = actions[:, : self.num_queries]
+ is_pad = is_pad[:, : self.num_queries]
+ num_noise_samples = self.noise_samples
+ # sample noise to add to actions
+ noise = torch.randn(
+ [num_noise_samples] + list(actions.shape), device=actions.device, dtype=actions.dtype
+ ) # num_noise, b, Ta, D(1, 2, 16, 14)
+ # sample a diffusion iteration for each data point
+ timesteps = torch.randint(
+ 0, self.noise_scheduler.config.num_train_timesteps, (b,), device=actions.device
+ ).long()
+
+ timesteps, noise = timesteps.to(actions.device), noise.to(actions.device)
+
+ # add noise to the clean actions according to the noise magnitude at each diffusion iteration
+ # (this is the forward diffusion process)
+ noisy_actions = torch.cat(
+ [self.noise_scheduler.add_noise(actions, noise[i], timesteps) for i in range(len(noise))],
+ dim=0,
+ ) # [num_noise_samples * b, Ta, action_dim]
+
+ noisy_actions = noisy_actions.to(dtype=actions.dtype)
+ assert hidden_states.ndim == 3
+
+ hidden_states = hidden_states.repeat(num_noise_samples, 1, 1)
+ timesteps = timesteps.repeat(num_noise_samples)
+ is_pad = is_pad.repeat(num_noise_samples, 1)
+ states = states.repeat(num_noise_samples, 1)
+
+ noise_pred = self.model_forward(
+ noisy_actions, timesteps, global_cond=hidden_states, states=states
+ )
+ noise = noise.view(noise.size(0) * noise.size(1), *noise.size()[2:])
+ loss = torch.nn.functional.mse_loss(noise_pred, noise, reduction="none")
+ loss = (loss * ~is_pad.unsqueeze(-1)).mean()
+ # loss_dict['loss'] = loss
+ return {"loss": loss}
+ # return loss
+ else: # inference time
+ b = 1
+ tp = self.num_queries
+ action_dim = self.action_dim
+
+ # initialize action from Gaussian noise
+ noisy_action = torch.randn((b, tp, action_dim)).cuda()
+
+ naction = noisy_action.to(dtype=hidden_states.dtype)
+ # init scheduler
+ self.noise_scheduler.set_timesteps(self.num_inference_timesteps)
+
+ for k in self.noise_scheduler.timesteps:
+ # predict noise
+ noise_pred = self.model_forward(naction, k, global_cond=hidden_states, states=states)
+
+ # inverse diffusion step (remove noise)
+ naction = self.noise_scheduler.step(
+ model_output=noise_pred, timestep=k, sample=naction
+ ).prev_sample
+
+ return naction
+
+ def model_forward(self, x, t, global_cond, states):
+ """
+ Forward pass of ScaleDP.
+ x: (N, T, input_dim) noisy actions
+ t: (N,) tensor of diffusion timesteps
+ global_cond: (N, n_obs_steps, D) tensor of conScaleDPions: image embeddings
+ """
+ global_cond = global_cond.squeeze(1)
+ global_cond = torch.cat([global_cond, states], dim=-1) if states is not None else global_cond
+ global_cond = self.combine(global_cond)
+
+ if not torch.is_tensor(t):
+ t = torch.tensor([t], dtype=torch.long, device=x.device)
+ elif torch.is_tensor(t) and len(t.shape) == 0:
+ t = t[None].to(x.device)
+ t = t.expand(t.shape[0])
+
+ x = self.x_embedder(x) + self.pos_embed.to(
+ device=x.device, dtype=x.dtype
+ ) # (N, T, D), where T = prediction_horizon
+ t = self.t_embedder(t) # (N, D)
+ if self.obs_as_cond:
+ global_cond = self.cond_obs_emb(global_cond) # (N, D)
+ # c = t + global_cond.sum(dim=1) # (N, D)
+ c = t + global_cond # (N, D)
+ for block in self.blocks:
+ # x = block(x, c, attn_mask=self.mask) # (N, T, D)
+ x = block(x, c, attn_mask=None) # (N, T, D)
+ x = self.final_layer(x, c) # (N, T, output_dim)
+ return x
+
+
+#################################################################################
+# Sine/Cosine Positional Embedding Functions #
+#################################################################################
+# https://github.com/facebookresearch/mae/blob/main/util/pos_embed.py
+
+
+def get_2d_sincos_pos_embed(embed_dim, grid_size, cls_token=False, extra_tokens=0):
+ """
+ grid_size: int of the grid height and width
+ return:
+ pos_embed: [grid_size*grid_size, embed_dim] or [1+grid_size*grid_size, embed_dim] (w/ or w/o cls_token)
+ """
+ grid_h = np.arange(grid_size, dtype=np.float32)
+ grid_w = np.arange(grid_size, dtype=np.float32)
+ grid = np.meshgrid(grid_w, grid_h) # here w goes first
+ grid = np.stack(grid, axis=0)
+
+ grid = grid.reshape([2, 1, grid_size, grid_size])
+ pos_embed = get_2d_sincos_pos_embed_from_grid(embed_dim, grid)
+ if cls_token and extra_tokens > 0:
+ pos_embed = np.concatenate([np.zeros([extra_tokens, embed_dim]), pos_embed], axis=0)
+ return pos_embed
+
+
+def get_2d_sincos_pos_embed_from_grid(embed_dim, grid):
+ assert embed_dim % 2 == 0
+
+ # use half of dimensions to encode grid_h
+ emb_h = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[0]) # (H*W, D/2)
+ emb_w = get_1d_sincos_pos_embed_from_grid(embed_dim // 2, grid[1]) # (H*W, D/2)
+
+ emb = np.concatenate([emb_h, emb_w], axis=1) # (H*W, D)
+ return emb
+
+
+def get_1d_sincos_pos_embed_from_grid(embed_dim, pos):
+ """
+ embed_dim: output dimension for each position
+ pos: a list of positions to be encoded: size (M,)
+ out: (M, D)
+ """
+ assert embed_dim % 2 == 0
+ omega = np.arange(embed_dim // 2, dtype=np.float64)
+ omega /= embed_dim / 2.0
+ omega = 1.0 / 10000**omega # (D/2,)
+
+ pos = pos.reshape(-1) # (M,)
+ out = np.einsum("m,d->md", pos, omega) # (M, D/2), outer product
+
+ emb_sin = np.sin(out) # (M, D/2)
+ emb_cos = np.cos(out) # (M, D/2)
+
+ emb = np.concatenate([emb_sin, emb_cos], axis=1) # (M, D)
+ return emb
+
+
+#################################################################################
+# ScaleDP Configs #
+#################################################################################
+
+
+def scaledp_h(**kwargs):
+ return ScaleDP(depth=32, n_emb=1280, num_heads=16, **kwargs)
+
+
+def scaledp_l(**kwargs):
+ return ScaleDP(depth=24, n_emb=1024, num_heads=16, **kwargs)
diff --git a/lerobot/common/policies/dexvla/policy_heads/modeling_unet_diffusion.py b/lerobot/common/policies/dexvla/policy_heads/modeling_unet_diffusion.py
new file mode 100644
index 00000000..0dea2e90
--- /dev/null
+++ b/lerobot/common/policies/dexvla/policy_heads/modeling_unet_diffusion.py
@@ -0,0 +1,387 @@
+#!/usr/bin/env python
+
+# Copyright 2025 DexVLA Team 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.
+
+import copy
+import math
+from typing import Union
+
+import torch
+import torch.nn as nn
+
+# requires diffusers==0.11.1
+from diffusers.schedulers.scheduling_ddim import DDIMScheduler
+from transformers.modeling_utils import PreTrainedModel
+
+from .configuration_unet_diffusion import UnetDiffusionPolicyConfig
+
+# =================== UNet for Diffusion ==============
+
+
+class SinusoidalPosEmb(nn.Module):
+ def __init__(self, dim, dtype):
+ super().__init__()
+ self.dim = dim
+ self.dtype = dtype
+
+ def forward(self, x):
+ device = x.device
+ half_dim = self.dim // 2
+ emb = math.log(10000) / (half_dim - 1)
+ emb = torch.exp(torch.arange(half_dim, device=device, dtype=self.dtype) * -emb)
+ emb = x[:, None] * emb[None, :]
+ emb = torch.cat((emb.sin(), emb.cos()), dim=-1)
+ return emb
+
+
+class Downsample1d(nn.Module):
+ def __init__(self, dim):
+ super().__init__()
+ self.conv = nn.Conv1d(dim, dim, 3, 2, 1)
+
+ def forward(self, x):
+ return self.conv(x)
+
+
+class Upsample1d(nn.Module):
+ def __init__(self, dim):
+ super().__init__()
+ self.conv = nn.ConvTranspose1d(dim, dim, 4, 2, 1)
+
+ def forward(self, x):
+ return self.conv(x)
+
+
+class Conv1dBlock(nn.Module):
+ """
+ Conv1d --> GroupNorm --> Mish
+ """
+
+ def __init__(self, inp_channels, out_channels, kernel_size, n_groups=8):
+ super().__init__()
+
+ self.block = nn.Sequential(
+ nn.Conv1d(inp_channels, out_channels, kernel_size, padding=kernel_size // 2),
+ nn.GroupNorm(n_groups, out_channels),
+ nn.Mish(),
+ )
+
+ def forward(self, x):
+ return self.block(x)
+
+
+class ConditionalResidualBlock1D(nn.Module):
+ def __init__(self, in_channels, out_channels, cond_dim, kernel_size=3, n_groups=8):
+ super().__init__()
+
+ self.blocks = nn.ModuleList(
+ [
+ Conv1dBlock(in_channels, out_channels, kernel_size, n_groups=n_groups),
+ Conv1dBlock(out_channels, out_channels, kernel_size, n_groups=n_groups),
+ ]
+ )
+
+ # FiLM modulation https://arxiv.org/abs/1709.07871
+ # predicts per-channel scale and bias
+ cond_channels = out_channels * 2
+ self.out_channels = out_channels
+ self.cond_encoder = nn.Sequential(
+ nn.Mish(), nn.Linear(cond_dim, cond_channels), nn.Unflatten(-1, (-1, 1))
+ )
+
+ # make sure dimensions compatible
+ self.residual_conv = (
+ nn.Conv1d(in_channels, out_channels, 1) if in_channels != out_channels else nn.Identity()
+ )
+
+ def forward(self, x, cond):
+ """
+ x : [ batch_size x in_channels x horizon ]
+ cond : [ batch_size x cond_dim]
+
+ returns:
+ out : [ batch_size x out_channels x horizon ]
+ """
+ out = self.blocks[0](x)
+ embed = self.cond_encoder(cond)
+
+ embed = embed.reshape(embed.shape[0], 2, self.out_channels, 1)
+ scale = embed[:, 0, ...]
+ bias = embed[:, 1, ...]
+ out = scale * out + bias
+
+ out = self.blocks[1](out)
+ out = out + self.residual_conv(x)
+ return out
+
+
+class ConditionalUnet1D(PreTrainedModel):
+ _no_split_modules = ["mid_modules", "down_modules", "up_modules"]
+
+ config_class = UnetDiffusionPolicyConfig
+
+ def __init__(self, config: UnetDiffusionPolicyConfig):
+ """
+ input_dim: Dim of actions.
+ global_cond_dim: Dim of global conditioning applied with FiLM
+ in addition to diffusion step embedding. This is usually obs_horizon * obs_dim
+ diffusion_step_embed_dim: Size of positional encoding for diffusion iteration k
+ down_dims: Channel size for each UNet level.
+ The length of this array determines number of levels.
+ kernel_size: Conv kernel size
+ n_groups: Number of groups for GroupNorm
+ """
+
+ super().__init__(config)
+ all_dims = [config.input_dim] + list(config.down_dims)
+ start_dim = config.down_dims[0]
+
+ self.num_queries = config.prediction_horizon
+ self.noise_samples = config.noise_samples
+ # self.global_1d_pool = nn.AdaptiveAvgPool1d(1)
+ # self.proj2action = nn.Linear(config.hidden_dim, config.global_cond_dim)
+ self.norm_after_pool = nn.LayerNorm(config.global_cond_dim)
+ self.combine = nn.Linear(config.global_cond_dim + config.state_dim, config.global_cond_dim)
+ dsed = config.diffusion_step_embed_dim
+ diffusion_step_encoder = nn.Sequential(
+ SinusoidalPosEmb(dsed, torch.bfloat16),
+ nn.Linear(dsed, dsed * 4),
+ nn.Mish(),
+ nn.Linear(dsed * 4, dsed),
+ )
+ cond_dim = dsed + config.global_cond_dim
+
+ in_out = list(zip(all_dims[:-1], all_dims[1:], strict=False))
+ mid_dim = all_dims[-1]
+ self.mid_modules = nn.ModuleList(
+ [
+ ConditionalResidualBlock1D(
+ mid_dim,
+ mid_dim,
+ cond_dim=cond_dim,
+ kernel_size=config.kernel_size,
+ n_groups=config.n_groups,
+ ),
+ ConditionalResidualBlock1D(
+ mid_dim,
+ mid_dim,
+ cond_dim=cond_dim,
+ kernel_size=config.kernel_size,
+ n_groups=config.n_groups,
+ ),
+ ]
+ )
+
+ down_modules = nn.ModuleList([])
+ for ind, (dim_in, dim_out) in enumerate(in_out):
+ is_last = ind >= (len(in_out) - 1)
+ down_modules.append(
+ nn.ModuleList(
+ [
+ ConditionalResidualBlock1D(
+ dim_in,
+ dim_out,
+ cond_dim=cond_dim,
+ kernel_size=config.kernel_size,
+ n_groups=config.n_groups,
+ ),
+ ConditionalResidualBlock1D(
+ dim_out,
+ dim_out,
+ cond_dim=cond_dim,
+ kernel_size=config.kernel_size,
+ n_groups=config.n_groups,
+ ),
+ Downsample1d(dim_out) if not is_last else nn.Identity(),
+ ]
+ )
+ )
+
+ up_modules = nn.ModuleList([])
+ for ind, (dim_in, dim_out) in enumerate(reversed(in_out[1:])):
+ is_last = ind >= (len(in_out) - 1)
+ up_modules.append(
+ nn.ModuleList(
+ [
+ ConditionalResidualBlock1D(
+ dim_out * 2,
+ dim_in,
+ cond_dim=cond_dim,
+ kernel_size=config.kernel_size,
+ n_groups=config.n_groups,
+ ),
+ ConditionalResidualBlock1D(
+ dim_in,
+ dim_in,
+ cond_dim=cond_dim,
+ kernel_size=config.kernel_size,
+ n_groups=config.n_groups,
+ ),
+ Upsample1d(dim_in) if not is_last else nn.Identity(),
+ ]
+ )
+ )
+
+ final_conv = nn.Sequential(
+ Conv1dBlock(start_dim, start_dim, kernel_size=config.kernel_size),
+ nn.Conv1d(start_dim, config.input_dim, 1),
+ )
+
+ self.diffusion_step_encoder = diffusion_step_encoder
+ self.up_modules = up_modules
+ self.down_modules = down_modules
+ self.final_conv = final_conv
+
+ print("number of parameters: {:e}".format(sum(p.numel() for p in self.parameters())))
+
+ self.num_inference_timesteps = config.num_inference_timesteps
+ # self.proj_to_action = nn.Identity()
+ self.noise_scheduler = DDIMScheduler(
+ num_train_timesteps=config.num_train_timesteps, # 100
+ beta_schedule="squaredcos_cap_v2",
+ clip_sample=True,
+ set_alpha_to_one=True,
+ steps_offset=0,
+ prediction_type="epsilon",
+ )
+
+ # self.num_inference_timesteps = config.num_inference_timesteps # 100
+
+ def forward(self, actions, hidden_states, states, is_pad):
+ """
+ Forward pass for the diffusion head.
+ :param actions: target actions, shape [b, Ta, D] D:10 = 3+6+1
+ :param hidden_states: hidden states from the llava_pythia, as the condition for the diffusion, shape [b,Tokens, D] 8 1200 1024
+ :param states: robot states, shape [b, D]
+ :return: loss
+ """
+ if actions is not None: # training time
+ b = actions.size(0)
+ actions = copy.deepcopy(actions[:, : self.num_queries])
+ is_pad = copy.deepcopy(is_pad[:, : self.num_queries])
+ num_noise_samples = self.noise_samples
+ # sample noise to add to actions
+ noise = torch.randn(
+ [num_noise_samples] + list(actions.shape), device=actions.device, dtype=actions.dtype
+ ) # num_noise, b, Ta, D
+ # sample a diffusion iteration for each data point
+ timesteps = torch.randint(
+ 0, self.noise_scheduler.config.num_train_timesteps, (b,), device=actions.device
+ ).long()
+
+ timesteps, noise = timesteps.to(actions.device), noise.to(actions.device)
+
+ # add noise to the clean actions according to the noise magnitude at each diffusion iteration
+ # (this is the forward diffusion process)
+ noisy_actions = torch.cat(
+ [self.noise_scheduler.add_noise(actions, noise[i], timesteps) for i in range(len(noise))],
+ dim=0,
+ ) # [num_noise_samples * b, Ta, action_dim]
+
+ noisy_actions = noisy_actions.to(dtype=actions.dtype)
+ assert hidden_states.ndim == 3
+
+ hidden_states = hidden_states.repeat(num_noise_samples, 1, 1)
+ timesteps = timesteps.repeat(num_noise_samples)
+ is_pad = is_pad.repeat(num_noise_samples, 1)
+ states = states.repeat(num_noise_samples, 1)
+
+ noise_pred = self.model_forward(
+ noisy_actions, timesteps, global_cond=hidden_states, states=states
+ )
+ noise = noise.view(noise.size(0) * noise.size(1), *noise.size()[2:])
+ loss = torch.nn.functional.mse_loss(noise_pred, noise, reduction="none")
+ loss = (loss * ~is_pad.unsqueeze(-1)).mean()
+ # loss_dict['loss'] = loss
+ return {"loss": loss}
+ # return loss
+ else: # inference time
+ b = 1
+ tp = self.num_queries
+ action_dim = 14
+
+ # initialize action from Gaussian noise
+ noisy_action = torch.randn((b, tp, action_dim)).cuda()
+
+ naction = noisy_action.to(dtype=hidden_states.dtype)
+ # init scheduler
+ self.noise_scheduler.set_timesteps(self.num_inference_timesteps)
+
+ for k in self.noise_scheduler.timesteps:
+ # predict noise
+ noise_pred = self.model_forward(naction, k, global_cond=hidden_states, states=states)
+
+ # inverse diffusion step (remove noise)
+ naction = self.noise_scheduler.step(
+ model_output=noise_pred, timestep=k, sample=naction
+ ).prev_sample
+
+ return naction
+
+ def model_forward(
+ self, sample: torch.Tensor, timestep: Union[torch.Tensor, float, int], global_cond=None, states=None
+ ):
+ """
+ x: (b,T,input_dim)
+ timestep: (b,) or int, diffusion step
+ global_cond: (b,global_cond_dim)
+ output: (b,T,input_dim)
+ """
+ # (b,t,c)
+ sample = sample.moveaxis(-1, -2)
+ # (b,c,t)
+ # global_cond = self.global_1d_pool(global_cond.permute(0, 2, 1)).squeeze(-1)
+ global_cond = global_cond.squeeze(1)
+
+ global_cond = self.norm_after_pool(global_cond)
+ global_cond = torch.cat([global_cond, states], dim=-1) if states is not None else global_cond
+ global_cond = self.combine(global_cond)
+ # 1. time
+ timesteps = timestep
+ if not torch.is_tensor(timesteps):
+ timesteps = torch.tensor([timesteps], dtype=torch.long, device=sample.device)
+ elif torch.is_tensor(timesteps) and len(timesteps.shape) == 0:
+ timesteps = timesteps[None].to(sample.device)
+ # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+ timesteps = timesteps.expand(sample.shape[0])
+
+ global_feature = self.diffusion_step_encoder(timesteps)
+
+ if global_cond is not None:
+ global_feature = torch.cat([global_feature, global_cond], axis=-1)
+
+ x = sample
+ h = []
+ for _idx, (resnet, resnet2, downsample) in enumerate(self.down_modules):
+ x = resnet(x, global_feature)
+ x = resnet2(x, global_feature)
+ h.append(x)
+ x = downsample(x)
+
+ for mid_module in self.mid_modules:
+ x = mid_module(x, global_feature)
+
+ for _idx, (resnet, resnet2, upsample) in enumerate(self.up_modules):
+ x = torch.cat((x, h.pop()), dim=1)
+ x = resnet(x, global_feature)
+ x = resnet2(x, global_feature)
+ x = upsample(x)
+
+ x = self.final_conv(x)
+
+ # (b,c,t)
+ x = x.moveaxis(-1, -2)
+ # (b,t,c)
+ return x
diff --git a/lerobot/common/policies/dexvla/qwe2_vla/__init__.py b/lerobot/common/policies/dexvla/qwe2_vla/__init__.py
new file mode 100644
index 00000000..35627635
--- /dev/null
+++ b/lerobot/common/policies/dexvla/qwe2_vla/__init__.py
@@ -0,0 +1,25 @@
+#!/usr/bin/env python
+
+# Copyright 2025 DexVLA Team 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.
+
+from transformers import AutoConfig, AutoModelForCausalLM
+
+from .configuration_qwen2_vla import Qwen2VLAConfig
+from .modeling_qwen2_vla import Qwen2VLForConditionalGenerationForVLA
+
+
+def register_qwen2_vla():
+ AutoConfig.register("qwen2_vla", Qwen2VLAConfig)
+ AutoModelForCausalLM.register(Qwen2VLAConfig, Qwen2VLForConditionalGenerationForVLA)
diff --git a/lerobot/common/policies/dexvla/qwe2_vla/configuration_qwen2_vla.py b/lerobot/common/policies/dexvla/qwe2_vla/configuration_qwen2_vla.py
new file mode 100644
index 00000000..1a3e7411
--- /dev/null
+++ b/lerobot/common/policies/dexvla/qwe2_vla/configuration_qwen2_vla.py
@@ -0,0 +1,254 @@
+#!/usr/bin/env python
+
+# Copyright 2025 DexVLA Team 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.
+
+import os
+from typing import Union
+
+from transformers.configuration_utils import PretrainedConfig
+from transformers.modeling_rope_utils import rope_config_validation
+from transformers.utils import logging
+
+logger = logging.get_logger(__name__)
+
+
+class Qwen2VLVisionConfig(PretrainedConfig):
+ model_type = "qwen2_vl"
+
+ def __init__(
+ self,
+ depth=32,
+ embed_dim=1280,
+ hidden_size=3584,
+ hidden_act="quick_gelu",
+ mlp_ratio=4,
+ num_heads=16,
+ in_channels=3,
+ patch_size=14,
+ spatial_merge_size=2,
+ temporal_patch_size=2,
+ **kwargs,
+ ):
+ super().__init__(**kwargs)
+
+ self.depth = depth
+ self.embed_dim = embed_dim
+ self.hidden_size = hidden_size
+ self.hidden_act = hidden_act
+ self.mlp_ratio = mlp_ratio
+ self.num_heads = num_heads
+ self.in_channels = in_channels
+ self.patch_size = patch_size
+ self.spatial_merge_size = spatial_merge_size
+ self.temporal_patch_size = temporal_patch_size
+
+ @classmethod
+ def from_pretrained(
+ cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs
+ ) -> "PretrainedConfig":
+ cls._set_token_in_kwargs(kwargs)
+
+ config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+ if config_dict.get("model_type") == "qwen2_vl":
+ config_dict = config_dict["vision_config"]
+
+ if (
+ "model_type" in config_dict
+ and hasattr(cls, "model_type")
+ and config_dict["model_type"] != cls.model_type
+ ):
+ logger.warning(
+ f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+ f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+ )
+
+ return cls.from_dict(config_dict, **kwargs)
+
+
+class Qwen2VLAConfig(PretrainedConfig):
+ r"""
+ This is the configuration class to store the configuration of a [`Qwen2VLModel`]. It is used to instantiate a
+ Qwen2-VL model according to the specified arguments, defining the model architecture. Instantiating a configuration
+ with the defaults will yield a similar configuration to that of
+ Qwen2-VL-7B-Instruct [Qwen/Qwen2-VL-7B-Instruct](https://huggingface.co/Qwen/Qwen2-VL-7B-Instruct).
+
+ Configuration objects inherit from [`PretrainedConfig`] and can be used to control the model outputs. Read the
+ documentation from [`PretrainedConfig`] for more information.
+
+
+ Args:
+ vocab_size (`int`, *optional*, defaults to 152064):
+ Vocabulary size of the Qwen2VL model. Defines the number of different tokens that can be represented by the
+ `inputs_ids` passed when calling [`Qwen2VLModel`]
+ hidden_size (`int`, *optional*, defaults to 8192):
+ Dimension of the hidden representations.
+ intermediate_size (`int`, *optional*, defaults to 29568):
+ Dimension of the MLP representations.
+ num_hidden_layers (`int`, *optional*, defaults to 80):
+ Number of hidden layers in the Transformer encoder.
+ num_attention_heads (`int`, *optional*, defaults to 64):
+ Number of attention heads for each attention layer in the Transformer encoder.
+ num_key_value_heads (`int`, *optional*, defaults to 8):
+ This is the number of key_value heads that should be used to implement Grouped Query Attention. If
+ `num_key_value_heads=num_attention_heads`, the model will use Multi Head Attention (MHA), if
+ `num_key_value_heads=1` the model will use Multi Query Attention (MQA) otherwise GQA is used. When
+ converting a multi-head checkpoint to a GQA checkpoint, each group key and value head should be constructed
+ by meanpooling all the original heads within that group. For more details checkout [this
+ paper](https://arxiv.org/pdf/2305.13245.pdf). If it is not specified, will default to `32`.
+ hidden_act (`str` or `function`, *optional*, defaults to `"silu"`):
+ The non-linear activation function (function or string) in the decoder.
+ max_position_embeddings (`int`, *optional*, defaults to 32768):
+ The maximum sequence length that this model might ever be used with.
+ initializer_range (`float`, *optional*, defaults to 0.02):
+ The standard deviation of the truncated_normal_initializer for initializing all weight matrices.
+ rms_norm_eps (`float`, *optional*, defaults to 1e-05):
+ The epsilon used by the rms normalization layers.
+ use_cache (`bool`, *optional*, defaults to `True`):
+ Whether or not the model should return the last key/values attentions (not used by all models). Only
+ relevant if `config.is_decoder=True`.
+ tie_word_embeddings (`bool`, *optional*, defaults to `False`):
+ Whether the model's input and output word embeddings should be tied.
+ rope_theta (`float`, *optional*, defaults to 1000000.0):
+ The base period of the RoPE embeddings.
+ use_sliding_window (`bool`, *optional*, defaults to `False`):
+ Whether to use sliding window attention.
+ sliding_window (`int`, *optional*, defaults to 4096):
+ Sliding window attention (SWA) window size. If not specified, will default to `4096`.
+ max_window_layers (`int`, *optional*, defaults to 80):
+ The number of layers that use SWA (Sliding Window Attention). The bottom layers use SWA while the top use full attention.
+ attention_dropout (`float`, *optional*, defaults to 0.0):
+ The dropout ratio for the attention probabilities.
+ vision_config (`Dict`, *optional*):
+ The config for the visual encoder initialization.
+ rope_scaling (`Dict`, *optional*):
+ Dictionary containing the scaling configuration for the RoPE embeddings. NOTE: if you apply new rope type
+ and you expect the model to work on longer `max_position_embeddings`, we recommend you to update this value
+ accordingly.
+ Expected contents:
+ `rope_type` (`str`):
+ The sub-variant of RoPE to use. Can be one of ['default', 'linear', 'dynamic', 'yarn', 'longrope',
+ 'llama3'], with 'default' being the original RoPE implementation.
+ `factor` (`float`, *optional*):
+ Used with all rope types except 'default'. The scaling factor to apply to the RoPE embeddings. In
+ most scaling types, a `factor` of x will enable the model to handle sequences of length x *
+ original maximum pre-trained length.
+ `original_max_position_embeddings` (`int`, *optional*):
+ Used with 'dynamic', 'longrope' and 'llama3'. The original max position embeddings used during
+ pretraining.
+ `attention_factor` (`float`, *optional*):
+ Used with 'yarn' and 'longrope'. The scaling factor to be applied on the attention
+ computation. If unspecified, it defaults to value recommended by the implementation, using the
+ `factor` field to infer the suggested value.
+ `beta_fast` (`float`, *optional*):
+ Only used with 'yarn'. Parameter to set the boundary for extrapolation (only) in the linear
+ ramp function. If unspecified, it defaults to 32.
+ `beta_slow` (`float`, *optional*):
+ Only used with 'yarn'. Parameter to set the boundary for interpolation (only) in the linear
+ ramp function. If unspecified, it defaults to 1.
+ `short_factor` (`List[float]`, *optional*):
+ Only used with 'longrope'. The scaling factor to be applied to short contexts (<
+ `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+ size divided by the number of attention heads divided by 2
+ `long_factor` (`List[float]`, *optional*):
+ Only used with 'longrope'. The scaling factor to be applied to long contexts (<
+ `original_max_position_embeddings`). Must be a list of numbers with the same length as the hidden
+ size divided by the number of attention heads divided by 2
+ `low_freq_factor` (`float`, *optional*):
+ Only used with 'llama3'. Scaling factor applied to low frequency components of the RoPE
+ `high_freq_factor` (`float`, *optional*):
+ Only used with 'llama3'. Scaling factor applied to high frequency components of the RoPE
+
+ ```python
+ >>> from transformers import Qwen2VLForConditionalGeneration, Qwen2VLConfig
+
+ >>> # Initializing a Qwen2VL style configuration
+ >>> configuration = Qwen2VLConfig()
+
+ >>> # Initializing a model from the Qwen2-VL-7B style configuration
+ >>> model = Qwen2VLForConditionalGeneration(configuration)
+
+ >>> # Accessing the model configuration
+ >>> configuration = model.config
+ ```"""
+
+ model_type = "qwen2_vla"
+ keys_to_ignore_at_inference = ["past_key_values"]
+
+ def __init__(
+ self,
+ vocab_size=152064,
+ hidden_size=8192,
+ intermediate_size=29568,
+ num_hidden_layers=80,
+ num_attention_heads=64,
+ num_key_value_heads=8,
+ hidden_act="silu",
+ max_position_embeddings=32768,
+ initializer_range=0.02,
+ rms_norm_eps=1e-05,
+ use_cache=True,
+ tie_word_embeddings=False,
+ rope_theta=1000000.0,
+ use_sliding_window=False,
+ sliding_window=4096,
+ max_window_layers=80,
+ attention_dropout=0.0,
+ vision_config=None,
+ rope_scaling=None,
+ # For loading policy head
+ policy_head_type="scale_dp_policy", # unet_diffusion_policy
+ **kwargs,
+ ):
+ if isinstance(vision_config, dict):
+ self.vision_config = Qwen2VLVisionConfig(**vision_config)
+ elif vision_config is None:
+ self.vision_config = Qwen2VLVisionConfig()
+
+ self.vocab_size = vocab_size
+ self.max_position_embeddings = max_position_embeddings
+ self.hidden_size = hidden_size
+ self.intermediate_size = intermediate_size
+ self.num_hidden_layers = num_hidden_layers
+ self.num_attention_heads = num_attention_heads
+ self.use_sliding_window = use_sliding_window
+ self.sliding_window = sliding_window
+ self.max_window_layers = max_window_layers
+ self.policy_head_type = policy_head_type # for loading policy head
+
+ # for backward compatibility
+ if num_key_value_heads is None:
+ num_key_value_heads = num_attention_heads
+
+ self.num_key_value_heads = num_key_value_heads
+ self.hidden_act = hidden_act
+ self.initializer_range = initializer_range
+ self.rms_norm_eps = rms_norm_eps
+ self.use_cache = use_cache
+ self.rope_theta = rope_theta
+ self.attention_dropout = attention_dropout
+ self.rope_scaling = rope_scaling
+
+ # Validate the correctness of rotary position embeddings parameters
+ # BC: if there is a 'type' field, move it to 'rope_type'.
+ # and change type from 'mrope' to 'default' because `mrope` does default RoPE calculations
+ # one can set it to "linear"/"dynamic" etc. to have scaled RoPE
+ if self.rope_scaling is not None and "type" in self.rope_scaling:
+ if self.rope_scaling["type"] == "mrope":
+ self.rope_scaling["type"] = "default"
+ self.rope_scaling["rope_type"] = self.rope_scaling["type"]
+ rope_config_validation(self, ignore_keys={"mrope_section"})
+
+ super().__init__(tie_word_embeddings=tie_word_embeddings, **kwargs)
diff --git a/lerobot/common/policies/dexvla/qwe2_vla/modeling_qwen2_vla.py b/lerobot/common/policies/dexvla/qwe2_vla/modeling_qwen2_vla.py
new file mode 100644
index 00000000..4b656354
--- /dev/null
+++ b/lerobot/common/policies/dexvla/qwe2_vla/modeling_qwen2_vla.py
@@ -0,0 +1,2046 @@
+#!/usr/bin/env python
+
+# Copyright 2025 DexVLA Team 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.
+"""PyTorch Qwen2-VL model."""
+
+import gc
+import math
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as func
+import torch.utils.checkpoint
+from torch.nn import CrossEntropyLoss, LayerNorm
+from transformers import AutoConfig, AutoModel
+from transformers.activations import ACT2FN
+from transformers.cache_utils import Cache, SlidingWindowCache, StaticCache
+from transformers.generation import GenerationMixin
+from transformers.modeling_attn_mask_utils import AttentionMaskConverter
+from transformers.modeling_outputs import (
+ BaseModelOutputWithPast,
+ ModelOutput,
+)
+from transformers.modeling_rope_utils import ROPE_INIT_FUNCTIONS
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import (
+ add_start_docstrings,
+ add_start_docstrings_to_model_forward,
+ is_flash_attn_2_available,
+ is_flash_attn_greater_or_equal_2_10,
+ logging,
+ replace_return_docstrings,
+)
+
+from lerobot.common.policies.dexvla.fusion_modules import ActionProjector, FiLM
+
+from .configuration_qwen2_vla import Qwen2VLAConfig, Qwen2VLVisionConfig
+
+if is_flash_attn_2_available():
+ from flash_attn import flash_attn_varlen_func
+ from transformers.modeling_flash_attention_utils import _flash_attention_forward
+else:
+ flash_attn_varlen_func = None
+
+
+logger = logging.get_logger(__name__)
+
+_CONFIG_FOR_DOC = "Qwen2VLConfig"
+
+
+@dataclass
+class Qwen2VLCausalLMOutputWithPast(ModelOutput):
+ """
+ Base class for Qwen2VL causal language model (or autoregressive) outputs.
+
+ Args:
+ loss (`torch.FloatTensor` of shape `(1,)`, *optional*, returned when `labels` is provided):
+ Language modeling loss (for next-token prediction).
+ logits (`torch.FloatTensor` of shape `(batch_size, sequence_length, config.vocab_size)`):
+ Prediction scores of the language modeling head (scores for each vocabulary token before SoftMax).
+ past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+ Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+ `(batch_size, num_heads, sequence_length, embed_size_per_head)`)
+
+ Contains pre-computed hidden-states (key and values in the self-attention blocks) that can be used (see
+ `past_key_values` input) to speed up sequential decoding.
+ hidden_states (`tuple(torch.FloatTensor)`, *optional*, returned when `output_hidden_states=True` is passed or when `config.output_hidden_states=True`):
+ Tuple of `torch.FloatTensor` (one for the output of the embeddings, if the model has an embedding layer, +
+ one for the output of each layer) of shape `(batch_size, sequence_length, hidden_size)`.
+
+ Hidden-states of the model at the output of each layer plus the optional initial embedding outputs.
+ attentions (`tuple(torch.FloatTensor)`, *optional*, returned when `output_attentions=True` is passed or when `config.output_attentions=True`):
+ Tuple of `torch.FloatTensor` (one for each layer) of shape `(batch_size, num_heads, sequence_length,
+ sequence_length)`.
+
+ Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+ heads.
+ rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
+ The rope index difference between sequence length and multimodal rope.
+ """
+
+ loss: Optional[torch.FloatTensor] = None
+ logits: torch.FloatTensor = None
+ past_key_values: Optional[List[torch.FloatTensor]] = None
+ hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+ attentions: Optional[Tuple[torch.FloatTensor]] = None
+ rope_deltas: Optional[torch.LongTensor] = None
+
+
+class Qwen2VLRotaryEmbedding(nn.Module):
+ def __init__(
+ self,
+ dim=None,
+ max_position_embeddings=2048,
+ base=10000,
+ device=None,
+ scaling_factor=1.0,
+ rope_type="default",
+ config: Optional[Qwen2VLAConfig] = None,
+ ):
+ super().__init__()
+ self.rope_kwargs = {}
+ if config is None:
+ logger.warning_once(
+ "`Qwen2VLRotaryEmbedding` can now be fully parameterized by passing the model config through the "
+ "`config` argument. All other arguments will be removed in v4.46"
+ )
+ self.rope_kwargs = {
+ "rope_type": rope_type,
+ "factor": scaling_factor,
+ "dim": dim,
+ "base": base,
+ "max_position_embeddings": max_position_embeddings,
+ }
+ self.rope_type = rope_type
+ self.max_seq_len_cached = max_position_embeddings
+ self.original_max_seq_len = max_position_embeddings
+ else:
+ # BC: "rope_type" was originally "type"
+ if config.rope_scaling is not None:
+ self.rope_type = config.rope_scaling.get("rope_type", config.rope_scaling.get("type"))
+ else:
+ self.rope_type = "default"
+ self.max_seq_len_cached = config.max_position_embeddings
+ self.original_max_seq_len = config.max_position_embeddings
+
+ self.config = config
+ self.rope_init_fn = ROPE_INIT_FUNCTIONS[self.rope_type]
+
+ inv_freq, self.attention_scaling = self.rope_init_fn(self.config, device, **self.rope_kwargs)
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
+ self.original_inv_freq = self.inv_freq
+
+ def _dynamic_frequency_update(self, position_ids, device):
+ """
+ dynamic RoPE layers should recompute `inv_freq` in the following situations:
+ 1 - growing beyond the cached sequence length (allow scaling)
+ 2 - the current sequence length is in the original scale (avoid losing precision with small sequences)
+ """
+ seq_len = torch.max(position_ids) + 1
+ if seq_len > self.max_seq_len_cached: # growth
+ inv_freq, self.attention_scaling = self.rope_init_fn(
+ self.config, device, seq_len=seq_len, **self.rope_kwargs
+ )
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
+ self.max_seq_len_cached = seq_len
+
+ if (
+ seq_len < self.original_max_seq_len and self.max_seq_len_cached > self.original_max_seq_len
+ ): # reset
+ self.register_buffer("inv_freq", self.original_inv_freq, persistent=False)
+ self.max_seq_len_cached = self.original_max_seq_len
+
+ @torch.no_grad()
+ def forward(self, x, position_ids):
+ if "dynamic" in self.rope_type:
+ self._dynamic_frequency_update(position_ids, device=x.device)
+
+ # Core RoPE block. In contrast to other models, Qwen2_VL has different position ids for spatiotemporal grids
+ # So we expand the inv_freq to shape (3, ...)
+ inv_freq_expanded = self.inv_freq[None, None, :, None].float().expand(3, position_ids.shape[1], -1, 1)
+ position_ids_expanded = position_ids[:, :, None, :].float() # shape (3, bs, 1, positions)
+ # Force float32 (see https://github.com/huggingface/transformers/pull/29285)
+ device_type = x.device.type
+ device_type = device_type if isinstance(device_type, str) and device_type != "mps" else "cpu"
+ with torch.autocast(device_type=device_type, enabled=False):
+ freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(2, 3)
+ emb = torch.cat((freqs, freqs), dim=-1)
+ cos = emb.cos()
+ sin = emb.sin()
+
+ # Advanced RoPE types (e.g. yarn) apply a post-processing scaling factor, equivalent to scaling attention
+ cos = cos * self.attention_scaling
+ sin = sin * self.attention_scaling
+
+ return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+ """Rotates half the hidden dims of the input."""
+ x1 = x[..., : x.shape[-1] // 2]
+ x2 = x[..., x.shape[-1] // 2 :]
+ return torch.cat((-x2, x1), dim=-1)
+
+
+def apply_multimodal_rotary_pos_emb(q, k, cos, sin, mrope_section, unsqueeze_dim=1):
+ """Applies Rotary Position Embedding with Multimodal Sections to the query and key tensors (https://qwenlm.github.io/blog/qwen2-vl/).
+
+ Explanation:
+ Multimodal 3D rotary position embedding is an extension to 1D rotary position embedding. The input embedding
+ sequence contains vision (images / videos) embedding and text embedding or just contains text embedding. For
+ vision embedding part, we apply rotary position embedding on temporal, height and width dimension separately.
+ Here we split the channel dimension to 3 chunks for the temporal, height and width rotary position embedding.
+ For text embedding part, we just apply 1D rotary position embedding. The three rotary position index (temporal,
+ height and width) of text embedding is always the same, so the text embedding rotary position embedding has no
+ difference with modern LLMs.
+
+ Args:
+ q (`torch.Tensor`): The query tensor.
+ k (`torch.Tensor`): The key tensor.
+ cos (`torch.Tensor`): The cosine part of the rotary embedding.
+ sin (`torch.Tensor`): The sine part of the rotary embedding.
+ position_ids (`torch.Tensor`):
+ The position indices of the tokens corresponding to the query and key tensors. For example, this can be
+ used to pass offsetted position ids when working with a KV-cache.
+ mrope_section(`List(int)`):
+ Multimodal rope section is for channel dimension of temporal, height and width in rope calculation.
+ unsqueeze_dim (`int`, *optional*, defaults to 1):
+ The 'unsqueeze_dim' argument specifies the dimension along which to unsqueeze cos[position_ids] and
+ sin[position_ids] so that they can be properly broadcasted to the dimensions of q and k. For example, note
+ that cos[position_ids] and sin[position_ids] have the shape [batch_size, seq_len, head_dim]. Then, if q and
+ k have the shape [batch_size, heads, seq_len, head_dim], then setting unsqueeze_dim=1 makes
+ cos[position_ids] and sin[position_ids] broadcastable to the shapes of q and k. Similarly, if q and k have
+ the shape [batch_size, seq_len, heads, head_dim], then set unsqueeze_dim=2.
+ Returns:
+ `tuple(torch.Tensor)` comprising of the query and key tensors rotated using the Rotary Position Embedding.
+ """
+ mrope_section = mrope_section * 2
+ cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
+ unsqueeze_dim
+ )
+ sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(
+ unsqueeze_dim
+ )
+
+ q_embed = (q * cos) + (rotate_half(q) * sin)
+ k_embed = (k * cos) + (rotate_half(k) * sin)
+ return q_embed, k_embed
+
+
+def apply_rotary_pos_emb_vision(tensor: torch.Tensor, freqs: torch.Tensor) -> torch.Tensor:
+ orig_dtype = tensor.dtype
+ tensor = tensor.float()
+ cos = freqs.cos()
+ sin = freqs.sin()
+ cos = cos.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
+ sin = sin.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
+ output = (tensor * cos) + (rotate_half(tensor) * sin)
+ output = output.to(orig_dtype)
+ return output
+
+
+class VisionRotaryEmbedding(nn.Module):
+ def __init__(self, dim: int, theta: float = 10000.0) -> None:
+ super().__init__()
+ inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
+ self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+ def forward(self, seqlen: int) -> torch.Tensor:
+ seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+ freqs = torch.outer(seq, self.inv_freq)
+ return freqs
+
+
+class PatchEmbed(nn.Module):
+ def __init__(
+ self,
+ patch_size: int = 14,
+ temporal_patch_size: int = 2,
+ in_channels: int = 3,
+ embed_dim: int = 1152,
+ ) -> None:
+ super().__init__()
+ self.patch_size = patch_size
+ self.temporal_patch_size = temporal_patch_size
+ self.in_channels = in_channels
+ self.embed_dim = embed_dim
+
+ kernel_size = [temporal_patch_size, patch_size, patch_size]
+ self.proj = nn.Conv3d(in_channels, embed_dim, kernel_size=kernel_size, stride=kernel_size, bias=False)
+
+ def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+ target_dtype = self.proj.weight.dtype
+ hidden_states = hidden_states.view(
+ -1, self.in_channels, self.temporal_patch_size, self.patch_size, self.patch_size
+ )
+ hidden_states = self.proj(hidden_states.to(dtype=target_dtype)).view(-1, self.embed_dim)
+ return hidden_states
+
+
+class PatchMerger(nn.Module):
+ def __init__(self, dim: int, context_dim: int, spatial_merge_size: int = 2) -> None:
+ super().__init__()
+ self.hidden_size = context_dim * (spatial_merge_size**2)
+ self.ln_q = LayerNorm(context_dim, eps=1e-6)
+ self.mlp = nn.Sequential(
+ nn.Linear(self.hidden_size, self.hidden_size),
+ nn.GELU(),
+ nn.Linear(self.hidden_size, dim),
+ )
+
+ def forward(self, x: torch.Tensor) -> torch.Tensor:
+ x = self.mlp(self.ln_q(x).view(-1, self.hidden_size))
+ return x
+
+
+class VisionMlp(nn.Module):
+ def __init__(self, dim: int, hidden_dim: int, hidden_act: str) -> None:
+ super().__init__()
+ self.fc1 = nn.Linear(dim, hidden_dim)
+ self.act = ACT2FN[hidden_act]
+ self.fc2 = nn.Linear(hidden_dim, dim)
+
+ def forward(self, x) -> torch.Tensor:
+ return self.fc2(self.act(self.fc1(x)))
+
+
+class VisionAttention(nn.Module):
+ def __init__(self, dim: int, num_heads: int = 16) -> None:
+ super().__init__()
+ self.num_heads = num_heads
+ self.head_dim = dim // num_heads
+ self.qkv = nn.Linear(dim, dim * 3, bias=True)
+ self.proj = nn.Linear(dim, dim)
+
+ def forward(
+ self, hidden_states: torch.Tensor, cu_seqlens: torch.Tensor, rotary_pos_emb: torch.Tensor = None
+ ) -> torch.Tensor:
+ seq_length = hidden_states.shape[0]
+ q, k, v = (
+ self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
+ )
+ q = apply_rotary_pos_emb_vision(q.unsqueeze(0), rotary_pos_emb).squeeze(0)
+ k = apply_rotary_pos_emb_vision(k.unsqueeze(0), rotary_pos_emb).squeeze(0)
+
+ attention_mask = torch.full(
+ [1, seq_length, seq_length], torch.finfo(q.dtype).min, device=q.device, dtype=q.dtype
+ )
+ for i in range(1, len(cu_seqlens)):
+ attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = 0
+
+ q = q.transpose(0, 1)
+ k = k.transpose(0, 1)
+ v = v.transpose(0, 1)
+ attn_weights = torch.matmul(q, k.transpose(1, 2)) / math.sqrt(self.head_dim)
+ attn_weights = attn_weights + attention_mask
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(q.dtype)
+ attn_output = torch.matmul(attn_weights, v)
+ attn_output = attn_output.transpose(0, 1)
+ attn_output = attn_output.reshape(seq_length, -1)
+ attn_output = self.proj(attn_output)
+ return attn_output
+
+
+class VisionFlashAttention2(nn.Module):
+ def __init__(self, dim: int, num_heads: int = 16) -> None:
+ super().__init__()
+ self.num_heads = num_heads
+ self.qkv = nn.Linear(dim, dim * 3, bias=True)
+ self.proj = nn.Linear(dim, dim)
+
+ def forward(
+ self, hidden_states: torch.Tensor, cu_seqlens: torch.Tensor, rotary_pos_emb: torch.Tensor = None
+ ) -> torch.Tensor:
+ seq_length = hidden_states.shape[0]
+ q, k, v = (
+ self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
+ )
+ q = apply_rotary_pos_emb_vision(q.unsqueeze(0), rotary_pos_emb).squeeze(0)
+ k = apply_rotary_pos_emb_vision(k.unsqueeze(0), rotary_pos_emb).squeeze(0)
+
+ max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
+ attn_output = flash_attn_varlen_func(q, k, v, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen).reshape(
+ seq_length, -1
+ )
+ attn_output = self.proj(attn_output)
+ return attn_output
+
+
+class VisionSdpaAttention(nn.Module):
+ def __init__(self, dim: int, num_heads: int = 16) -> None:
+ super().__init__()
+ self.num_heads = num_heads
+ self.qkv = nn.Linear(dim, dim * 3, bias=True)
+ self.proj = nn.Linear(dim, dim)
+
+ def forward(
+ self, hidden_states: torch.Tensor, cu_seqlens: torch.Tensor, rotary_pos_emb: torch.Tensor = None
+ ) -> torch.Tensor:
+ seq_length = hidden_states.shape[0]
+ q, k, v = (
+ self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
+ )
+ q = apply_rotary_pos_emb_vision(q.unsqueeze(0), rotary_pos_emb).squeeze(0)
+ k = apply_rotary_pos_emb_vision(k.unsqueeze(0), rotary_pos_emb).squeeze(0)
+
+ attention_mask = torch.zeros([1, seq_length, seq_length], device=q.device, dtype=torch.bool)
+ for i in range(1, len(cu_seqlens)):
+ attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = True
+ q = q.transpose(0, 1)
+ k = k.transpose(0, 1)
+ v = v.transpose(0, 1)
+ attn_output = func.scaled_dot_product_attention(q, k, v, attention_mask, dropout_p=0.0)
+ attn_output = attn_output.transpose(0, 1)
+ attn_output = attn_output.reshape(seq_length, -1)
+ attn_output = self.proj(attn_output)
+ return attn_output
+
+
+QWEN2_VL_VISION_ATTENTION_CLASSES = {
+ "eager": VisionAttention,
+ "flash_attention_2": VisionFlashAttention2,
+ "sdpa": VisionSdpaAttention,
+}
+
+
+class Qwen2VLVisionBlock(nn.Module):
+ def __init__(self, config, attn_implementation: str = "sdpa") -> None:
+ super().__init__()
+ self.norm1 = LayerNorm(config.embed_dim, eps=1e-6)
+ self.norm2 = LayerNorm(config.embed_dim, eps=1e-6)
+ mlp_hidden_dim = int(config.embed_dim * config.mlp_ratio)
+
+ self.attn = QWEN2_VL_VISION_ATTENTION_CLASSES[attn_implementation](
+ config.embed_dim, num_heads=config.num_heads
+ )
+ self.mlp = VisionMlp(dim=config.embed_dim, hidden_dim=mlp_hidden_dim, hidden_act=config.hidden_act)
+
+ def forward(self, hidden_states, cu_seqlens, rotary_pos_emb) -> torch.Tensor:
+ hidden_states = hidden_states + self.attn(
+ self.norm1(hidden_states).to(torch.bfloat16), cu_seqlens=cu_seqlens, rotary_pos_emb=rotary_pos_emb
+ )
+ hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
+ return hidden_states
+
+
+# Copied from transformers.models.qwen2.modeling_qwen2.Qwen2RMSNorm
+class Qwen2RMSNorm(nn.Module):
+ def __init__(self, hidden_size, eps=1e-6):
+ """
+ Qwen2RMSNorm is equivalent to T5LayerNorm
+ """
+ super().__init__()
+ self.weight = nn.Parameter(torch.ones(hidden_size))
+ self.variance_epsilon = eps
+
+ def forward(self, hidden_states):
+ input_dtype = hidden_states.dtype
+ hidden_states = hidden_states.to(torch.float32)
+ variance = hidden_states.pow(2).mean(-1, keepdim=True)
+ hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
+ return self.weight * hidden_states.to(input_dtype)
+
+ def extra_repr(self):
+ return f"{tuple(self.weight.shape)}, eps={self.variance_epsilon}"
+
+
+# Copied from transformers.models.qwen2.modeling_qwen2.Qwen2MLP
+class Qwen2MLP(nn.Module):
+ def __init__(self, config):
+ super().__init__()
+ self.hidden_size = config.hidden_size
+ self.intermediate_size = config.intermediate_size
+ self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+ self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
+ self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
+ self.act_fn = ACT2FN[config.hidden_act]
+
+ def forward(self, hidden_state):
+ return self.down_proj(self.act_fn(self.gate_proj(hidden_state)) * self.up_proj(hidden_state))
+
+
+# Copied from transformers.models.llama.modeling_llama.repeat_kv
+def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
+ """
+ This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
+ num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
+ """
+ batch, num_key_value_heads, slen, head_dim = hidden_states.shape
+ if n_rep == 1:
+ return hidden_states
+ hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
+ return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)
+
+
+class Qwen2VLAttention(nn.Module):
+ """
+ Multi-headed attention from 'Attention Is All You Need' paper. Modified to use sliding window attention: Longformer
+ and "Generating Long Sequences with Sparse Transformers".
+ """
+
+ def __init__(self, config: Qwen2VLAConfig, layer_idx: Optional[int] = None):
+ super().__init__()
+ self.config = config
+ self.layer_idx = layer_idx
+ if layer_idx is None:
+ logger.warning_once(
+ f"Instantiating {self.__class__.__name__} without passing `layer_idx` is not recommended and will "
+ "to errors during the forward call, if caching is used. Please make sure to provide a `layer_idx` "
+ "when creating this class."
+ )
+
+ self.hidden_size = config.hidden_size
+ self.num_heads = config.num_attention_heads
+ self.head_dim = self.hidden_size // self.num_heads
+ self.num_key_value_heads = config.num_key_value_heads
+ self.num_key_value_groups = self.num_heads // self.num_key_value_heads
+ self.max_position_embeddings = config.max_position_embeddings
+ self.rope_theta = config.rope_theta
+ self.is_causal = True
+ self.attention_dropout = config.attention_dropout
+ self.rope_scaling = config.rope_scaling
+
+ if (self.head_dim * self.num_heads) != self.hidden_size:
+ raise ValueError(
+ f"hidden_size must be divisible by num_heads (got `hidden_size`: {self.hidden_size}"
+ f" and `num_heads`: {self.num_heads})."
+ )
+ self.q_proj = nn.Linear(self.hidden_size, self.num_heads * self.head_dim, bias=True)
+ self.k_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
+ self.v_proj = nn.Linear(self.hidden_size, self.num_key_value_heads * self.head_dim, bias=True)
+ self.o_proj = nn.Linear(self.num_heads * self.head_dim, self.hidden_size, bias=False)
+
+ self.rotary_emb = Qwen2VLRotaryEmbedding(
+ self.head_dim,
+ max_position_embeddings=self.max_position_embeddings,
+ base=self.rope_theta,
+ )
+
+ def forward(
+ self,
+ hidden_states: torch.Tensor,
+ attention_mask: Optional[torch.Tensor] = None,
+ position_ids: Optional[torch.LongTensor] = None,
+ past_key_value: Optional[Cache] = None,
+ output_attentions: bool = False,
+ use_cache: bool = False,
+ cache_position: Optional[torch.LongTensor] = None,
+ position_embeddings: Optional[
+ Tuple[torch.Tensor, torch.Tensor]
+ ] = None, # will become mandatory in v4.46
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+ bsz, q_len, _ = hidden_states.size()
+
+ query_states = self.q_proj(hidden_states)
+ key_states = self.k_proj(hidden_states)
+ value_states = self.v_proj(hidden_states)
+
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+ kv_seq_len = key_states.shape[-2]
+ if past_key_value is not None:
+ kv_seq_len += cache_position[0] + 1
+
+ if position_embeddings is None:
+ logger.warning_once(
+ "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+ "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
+ "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
+ "removed and `position_embeddings` will be mandatory."
+ )
+ cos, sin = self.rotary_emb(value_states, position_ids)
+ else:
+ cos, sin = position_embeddings
+ query_states, key_states = apply_multimodal_rotary_pos_emb(
+ query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
+ )
+
+ if past_key_value is not None:
+ cache_kwargs = {
+ "sin": sin,
+ "cos": cos,
+ "cache_position": cache_position,
+ } # Specific to RoPE models
+ key_states, value_states = past_key_value.update(
+ key_states, value_states, self.layer_idx, cache_kwargs
+ )
+
+ # repeat k/v heads if n_kv_heads < n_heads
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+ attn_weights = torch.matmul(query_states, key_states.transpose(2, 3)) / math.sqrt(self.head_dim)
+
+ if attention_mask is not None: # no matter the length, we just slice it
+ causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+ attn_weights = attn_weights + causal_mask
+
+ # Fix precision issues in Qwen2-VL float16 inference
+ # Replace inf values with zeros in attention weights to prevent NaN propagation
+ if query_states.dtype == torch.float16:
+ attn_weights = torch.where(
+ torch.isinf(attn_weights), torch.zeros_like(attn_weights), attn_weights
+ )
+
+ # upcast attention to fp32
+ attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+ attn_weights = nn.functional.dropout(attn_weights, p=self.attention_dropout, training=self.training)
+ attn_output = torch.matmul(attn_weights, value_states)
+
+ if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
+ raise ValueError(
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
+ f" {attn_output.size()}"
+ )
+
+ attn_output = attn_output.transpose(1, 2).contiguous()
+ attn_output = attn_output.reshape(bsz, q_len, -1)
+
+ attn_output = self.o_proj(attn_output)
+
+ if not output_attentions:
+ attn_weights = None
+
+ return attn_output, attn_weights, past_key_value
+
+
+class Qwen2VLFlashAttention2(Qwen2VLAttention):
+ """
+ Qwen2VL flash attention module, following Qwen2VL attention module. This module inherits from `Qwen2VLAttention`
+ as the weights of the module stays untouched. The only required change would be on the forward pass
+ where it needs to correctly call the public API of flash attention and deal with padding tokens
+ in case the input contains any of them. Additionally, for sliding window attention, we apply SWA only to the bottom
+ config.max_window_layers layers.
+ """
+
+ def __init__(self, *args, **kwargs):
+ super().__init__(*args, **kwargs)
+
+ # flash_attn<2.1 generates top-left aligned causal mask, while what is needed here is bottom-right alignment, that was made default for flash_attn>=2.1. This attribute is used to handle this difference. Reference: https://github.com/Dao-AILab/flash-attention/releases/tag/v2.1.0.
+ # Beware that with flash_attn<2.1, using q_seqlen != k_seqlen (except for the case q_seqlen == 1) produces a wrong mask (top-left).
+ self._flash_attn_uses_top_left_mask = not is_flash_attn_greater_or_equal_2_10()
+
+ def forward(
+ self,
+ hidden_states: torch.Tensor,
+ attention_mask: Optional[torch.Tensor] = None,
+ position_ids: Optional[torch.LongTensor] = None,
+ past_key_value: Optional[Cache] = None,
+ output_attentions: bool = False,
+ use_cache: bool = False,
+ cache_position: Optional[torch.LongTensor] = None,
+ position_embeddings: Optional[
+ Tuple[torch.Tensor, torch.Tensor]
+ ] = None, # will become mandatory in v4.46
+ ):
+ bsz, q_len, _ = hidden_states.size()
+
+ query_states = self.q_proj(hidden_states)
+ key_states = self.k_proj(hidden_states)
+ value_states = self.v_proj(hidden_states)
+
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+ kv_seq_len = key_states.shape[-2]
+ if past_key_value is not None:
+ if self.layer_idx is None:
+ raise ValueError(
+ f"The cache structure has changed since version v4.36. If you are using {self.__class__.__name__} "
+ "for auto-regressive decoding with k/v caching, please make sure to initialize the attention class "
+ "with a layer index."
+ )
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+
+ # Because the input can be padded, the absolute sequence length depends on the max position id.
+ if position_embeddings is None:
+ logger.warning_once(
+ "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+ "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
+ "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
+ "removed and `position_embeddings` will be mandatory."
+ )
+ cos, sin = self.rotary_emb(value_states, position_ids)
+ else:
+ cos, sin = position_embeddings
+
+ query_states, key_states = apply_multimodal_rotary_pos_emb(
+ query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
+ )
+
+ if past_key_value is not None:
+ # Activate slicing cache only if the config has a value `sliding_windows` attribute
+ cache_has_contents = past_key_value.get_seq_length(self.layer_idx) > 0
+ if (
+ getattr(self.config, "sliding_window", None) is not None
+ and kv_seq_len > self.config.sliding_window
+ and cache_has_contents
+ ):
+ slicing_tokens = 1 - self.config.sliding_window
+
+ past_key = past_key_value[self.layer_idx][0]
+ past_value = past_key_value[self.layer_idx][1]
+
+ past_key = past_key[:, :, slicing_tokens:, :].contiguous()
+ past_value = past_value[:, :, slicing_tokens:, :].contiguous()
+
+ if past_key.shape[-2] != self.config.sliding_window - 1:
+ raise ValueError(
+ f"past key must have a shape of (`batch_size, num_heads, self.config.sliding_window-1, head_dim`), got"
+ f" {past_key.shape}"
+ )
+
+ if attention_mask is not None:
+ attention_mask = attention_mask[:, slicing_tokens:]
+ attention_mask = torch.cat(
+ [attention_mask, torch.ones_like(attention_mask[:, -1:])], dim=-1
+ )
+
+ cache_kwargs = {
+ "sin": sin,
+ "cos": cos,
+ "cache_position": cache_position,
+ } # Specific to RoPE models
+ key_states, value_states = past_key_value.update(
+ key_states, value_states, self.layer_idx, cache_kwargs
+ )
+
+ # repeat k/v heads if n_kv_heads < n_heads
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
+ dropout_rate = 0.0 if not self.training else self.attention_dropout
+
+ # In PEFT, usually we cast the layer norms in float32 for training stability reasons
+ # therefore the input hidden states gets silently casted in float32. Hence, we need
+ # cast them back in float16 just to be sure everything works as expected.
+ input_dtype = query_states.dtype
+ if input_dtype == torch.float32:
+ if torch.is_autocast_enabled():
+ target_dtype = torch.get_autocast_gpu_dtype()
+ # Handle the case where the model is quantized
+ elif hasattr(self.config, "_pre_quantization_dtype"):
+ target_dtype = self.config._pre_quantization_dtype
+ else:
+ target_dtype = self.q_proj.weight.dtype
+
+ logger.warning_once(
+ f"The input hidden states seems to be silently casted in float32, this might be related to"
+ f" the fact you have upcasted embedding or layer norm layers in float32. We will cast back the input in"
+ f" {target_dtype}."
+ )
+
+ query_states = query_states.to(target_dtype)
+ key_states = key_states.to(target_dtype)
+ value_states = value_states.to(target_dtype)
+
+ # Reashape to the expected shape for Flash Attention
+ query_states = query_states.transpose(1, 2)
+ key_states = key_states.transpose(1, 2)
+ value_states = value_states.transpose(1, 2)
+
+ if (
+ self.config.use_sliding_window
+ and getattr(self.config, "sliding_window", None) is not None
+ and self.layer_idx >= self.config.max_window_layers
+ ):
+ sliding_window = self.config.sliding_window
+ else:
+ sliding_window = None
+
+ attn_output = _flash_attention_forward(
+ query_states,
+ key_states,
+ value_states,
+ attention_mask,
+ q_len,
+ dropout=dropout_rate,
+ sliding_window=sliding_window,
+ is_causal=self.is_causal,
+ use_top_left_mask=self._flash_attn_uses_top_left_mask,
+ )
+
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size).contiguous()
+ attn_output = self.o_proj(attn_output)
+
+ if not output_attentions:
+ attn_weights = None
+
+ return attn_output, attn_weights, past_key_value
+
+
+class Qwen2VLSdpaAttention(Qwen2VLAttention):
+ """
+ Qwen2 attention module using torch.nn.functional.scaled_dot_product_attention. This module inherits from
+ `Qwen2Attention` as the weights of the module stays untouched. The only changes are on the forward pass to adapt to
+ SDPA API.
+ """
+
+ # Adapted from Qwen2Attention.forward
+ def forward(
+ self,
+ hidden_states: torch.Tensor,
+ attention_mask: Optional[torch.Tensor] = None,
+ position_ids: Optional[torch.LongTensor] = None,
+ past_key_value: Optional[Cache] = None,
+ output_attentions: bool = False,
+ use_cache: bool = False,
+ cache_position: Optional[torch.LongTensor] = None,
+ position_embeddings: Optional[
+ Tuple[torch.Tensor, torch.Tensor]
+ ] = None, # will become mandatory in v4.46
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
+ if output_attentions:
+ logger.warning_once(
+ "Qwen2VLModel is using Qwen2VLSdpaAttention, but `torch.nn.functional.scaled_dot_product_attention` does not support `output_attentions=True`. Falling back to the manual attention implementation, "
+ 'but specifying the manual implementation will be required from Transformers version v5.0.0 onwards. This warning can be removed using the argument `attn_implementation="eager"` when loading the model.'
+ )
+ return super().forward(
+ hidden_states=hidden_states,
+ attention_mask=attention_mask,
+ position_ids=position_ids,
+ past_key_value=past_key_value,
+ output_attentions=output_attentions,
+ use_cache=use_cache,
+ cache_position=cache_position,
+ )
+
+ bsz, q_len, _ = hidden_states.size()
+
+ query_states = self.q_proj(hidden_states)
+ key_states = self.k_proj(hidden_states)
+ value_states = self.v_proj(hidden_states)
+
+ query_states = query_states.view(bsz, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+ key_states = key_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+ value_states = value_states.view(bsz, q_len, self.num_key_value_heads, self.head_dim).transpose(1, 2)
+
+ kv_seq_len = key_states.shape[-2]
+ if past_key_value is not None:
+ kv_seq_len += past_key_value.get_usable_length(kv_seq_len, self.layer_idx)
+ if position_embeddings is None:
+ logger.warning_once(
+ "The attention layers in this model are transitioning from computing the RoPE embeddings internally "
+ "through `position_ids` (2D tensor with the indexes of the tokens), to using externally computed "
+ "`position_embeddings` (Tuple of tensors, containing cos and sin). In v4.46 `position_ids` will be "
+ "removed and `position_embeddings` will be mandatory."
+ )
+ cos, sin = self.rotary_emb(value_states, position_ids)
+ else:
+ cos, sin = position_embeddings
+ query_states, key_states = apply_multimodal_rotary_pos_emb(
+ query_states, key_states, cos, sin, self.rope_scaling["mrope_section"]
+ )
+
+ if past_key_value is not None:
+ cache_kwargs = {
+ "sin": sin,
+ "cos": cos,
+ "cache_position": cache_position,
+ } # Specific to RoPE models
+ key_states, value_states = past_key_value.update(
+ key_states, value_states, self.layer_idx, cache_kwargs
+ )
+
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
+
+ causal_mask = attention_mask
+ if attention_mask is not None: # no matter the length, we just slice it
+ causal_mask = attention_mask[:, :, :, : key_states.shape[-2]]
+
+ # SDPA with memory-efficient backend is currently (torch==2.1.2) bugged with non-contiguous inputs with custom attn_mask,
+ # Reference: https://github.com/pytorch/pytorch/issues/112577.
+ if query_states.device.type == "cuda" and attention_mask is not None:
+ query_states = query_states.contiguous()
+ key_states = key_states.contiguous()
+ value_states = value_states.contiguous()
+
+ # We dispatch to SDPA's Flash Attention or Efficient kernels via this `is_causal` if statement instead of an inline conditional assignment
+ # in SDPA to support both torch.compile's dynamic shapes and full graph options. An inline conditional prevents dynamic shapes from compiling.
+ # The q_len > 1 is necessary to match with AttentionMaskConverter.to_causal_4d that does not create a causal mask in case q_len == 1.
+ is_causal = bool(causal_mask is None and q_len > 1)
+
+ attn_output = torch.nn.functional.scaled_dot_product_attention(
+ query_states,
+ key_states,
+ value_states,
+ attn_mask=causal_mask,
+ dropout_p=self.attention_dropout if self.training else 0.0,
+ is_causal=is_causal,
+ )
+
+ attn_output = attn_output.transpose(1, 2).contiguous()
+ attn_output = attn_output.view(bsz, q_len, self.hidden_size)
+
+ attn_output = self.o_proj(attn_output)
+
+ return attn_output, None, past_key_value
+
+
+QWEN2_VL_ATTENTION_CLASSES = {
+ "eager": Qwen2VLAttention,
+ "flash_attention_2": Qwen2VLFlashAttention2,
+ "sdpa": Qwen2VLSdpaAttention,
+}
+
+
+class Qwen2VLDecoderLayer(nn.Module):
+ def __init__(self, config: Qwen2VLAConfig, layer_idx: int):
+ super().__init__()
+ self.hidden_size = config.hidden_size
+
+ if config.use_sliding_window and config._attn_implementation != "flash_attention_2":
+ logger.warning_once(
+ f"Sliding Window Attention is enabled but not implemented for `{config._attn_implementation}`; "
+ "unexpected results may be encountered."
+ )
+ self.self_attn = QWEN2_VL_ATTENTION_CLASSES[config._attn_implementation](config, layer_idx)
+
+ self.mlp = Qwen2MLP(config)
+ self.input_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+ self.post_attention_layernorm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+ def forward(
+ self,
+ hidden_states: torch.Tensor,
+ attention_mask: Optional[torch.Tensor] = None,
+ position_ids: Optional[torch.LongTensor] = None,
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
+ output_attentions: Optional[bool] = False,
+ use_cache: Optional[bool] = False,
+ cache_position: Optional[torch.LongTensor] = None,
+ position_embeddings: Optional[
+ Tuple[torch.Tensor, torch.Tensor]
+ ] = None, # will become mandatory in v4.46
+ **kwargs,
+ ) -> Tuple[torch.FloatTensor, Optional[Tuple[torch.FloatTensor, torch.FloatTensor]]]:
+ """
+ Args:
+ hidden_states (`torch.FloatTensor`): input to the layer of shape `(batch, seq_len, embed_dim)`
+ attention_mask (`torch.FloatTensor`, *optional*): attention mask of size
+ `(batch, sequence_length)` where padding elements are indicated by 0.
+ output_attentions (`bool`, *optional*):
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under
+ returned tensors for more detail.
+ use_cache (`bool`, *optional*):
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding
+ (see `past_key_values`).
+ past_key_value (`Tuple(torch.FloatTensor)`, *optional*): cached past key and value projection states
+ cache_position (`torch.LongTensor` of shape `(sequence_length)`, *optional*):
+ Indices depicting the position of the input sequence tokens in the sequence.
+ position_embeddings (`Tuple[torch.FloatTensor, torch.FloatTensor]`, *optional*):
+ Tuple containing the cosine and sine positional embeddings of shape `(batch_size, seq_len, head_dim)`,
+ with `head_dim` being the embedding dimension of each attention head.
+ kwargs (`dict`, *optional*):
+ Arbitrary kwargs to be ignored, used for FSDP and other methods that injects code
+ into the model
+ """
+
+ residual = hidden_states
+
+ hidden_states = self.input_layernorm(hidden_states)
+
+ # Self Attention
+ hidden_states, self_attn_weights, present_key_value = self.self_attn(
+ hidden_states=hidden_states,
+ attention_mask=attention_mask,
+ position_ids=position_ids,
+ past_key_value=past_key_value,
+ output_attentions=output_attentions,
+ use_cache=use_cache,
+ cache_position=cache_position,
+ position_embeddings=position_embeddings,
+ )
+ hidden_states = residual + hidden_states
+
+ # Fully Connected
+ residual = hidden_states
+ hidden_states = self.post_attention_layernorm(hidden_states)
+ hidden_states = self.mlp(hidden_states)
+ hidden_states = residual + hidden_states
+
+ outputs = (hidden_states,)
+
+ if output_attentions:
+ outputs += (self_attn_weights,)
+
+ if use_cache:
+ outputs += (present_key_value,)
+
+ return outputs
+
+
+QWEN2VL_START_DOCSTRING = r"""
+ This model inherits from [`PreTrainedModel`]. Check the superclass documentation for the generic methods the
+ library implements for all its model (such as downloading or saving, resizing the input embeddings, pruning heads
+ etc.)
+
+ This model is also a PyTorch [torch.nn.Module](https://pytorch.org/docs/stable/nn.html#torch.nn.Module) subclass.
+ Use it as a regular PyTorch Module and refer to the PyTorch documentation for all matter related to general usage
+ and behavior.
+
+ Parameters:
+ config ([`Qwen2VLConfig`]):
+ Model configuration class with all the parameters of the model. Initializing with a config file does not
+ load the weights associated with the model, only the configuration. Check out the
+ [`~PreTrainedModel.from_pretrained`] method to load the model weights.
+"""
+
+
+@add_start_docstrings(
+ "The bare Qwen2VL Model outputting raw hidden-states without any specific head on top.",
+ QWEN2VL_START_DOCSTRING,
+)
+class Qwen2VLPreTrainedModel(PreTrainedModel):
+ config_class = Qwen2VLAConfig
+ base_model_prefix = "model"
+ supports_gradient_checkpointing = True
+ _no_split_modules = ["Qwen2VLDecoderLayer", "Qwen2VLVisionBlock", "policy_head"]
+ _skip_keys_device_placement = "past_key_values"
+ _supports_flash_attn_2 = True
+ _supports_sdpa = True
+ _supports_cache_class = True
+ _supports_static_cache = True
+
+ def _init_weights(self, module):
+ std = self.config.initializer_range
+ if isinstance(module, (nn.Linear, nn.Conv3d)):
+ module.weight.data.normal_(mean=0.0, std=std)
+ if module.bias is not None:
+ module.bias.data.zero_()
+ elif isinstance(module, nn.Embedding):
+ module.weight.data.normal_(mean=0.0, std=std)
+ if module.padding_idx is not None:
+ module.weight.data[module.padding_idx].zero_()
+
+
+class Qwen2VisionTransformerPretrainedModel(Qwen2VLPreTrainedModel):
+ config_class = Qwen2VLVisionConfig
+ _no_split_modules = ["Qwen2VLVisionBlock"]
+
+ def __init__(self, config) -> None:
+ super().__init__(config)
+ self.spatial_merge_size = config.spatial_merge_size
+
+ self.patch_embed = PatchEmbed(
+ patch_size=config.patch_size,
+ temporal_patch_size=config.temporal_patch_size,
+ in_channels=config.in_channels,
+ embed_dim=config.embed_dim,
+ )
+
+ head_dim = config.embed_dim // config.num_heads
+ self.rotary_pos_emb = VisionRotaryEmbedding(head_dim // 2)
+
+ self.blocks = nn.ModuleList(
+ [Qwen2VLVisionBlock(config, config._attn_implementation) for _ in range(config.depth)]
+ )
+ self.merger = PatchMerger(
+ dim=config.hidden_size, context_dim=config.embed_dim, spatial_merge_size=config.spatial_merge_size
+ )
+
+ def get_dtype(self) -> torch.dtype:
+ return self.blocks[0].mlp.fc2.weight.dtype
+
+ def get_device(self) -> torch.device:
+ return self.blocks[0].mlp.fc2.weight.device
+
+ def rot_pos_emb(self, grid_spatiotemporal):
+ pos_ids = []
+ for t, h, w in grid_spatiotemporal:
+ hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
+ hpos_ids = hpos_ids.reshape(
+ h // self.spatial_merge_size,
+ self.spatial_merge_size,
+ w // self.spatial_merge_size,
+ self.spatial_merge_size,
+ )
+ hpos_ids = hpos_ids.permute(0, 2, 1, 3)
+ hpos_ids = hpos_ids.flatten()
+
+ wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
+ wpos_ids = wpos_ids.reshape(
+ h // self.spatial_merge_size,
+ self.spatial_merge_size,
+ w // self.spatial_merge_size,
+ self.spatial_merge_size,
+ )
+ wpos_ids = wpos_ids.permute(0, 2, 1, 3)
+ wpos_ids = wpos_ids.flatten()
+ pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
+ pos_ids = torch.cat(pos_ids, dim=0)
+ max_grid_size = grid_spatiotemporal[:, 1:].max()
+ rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+ rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
+ return rotary_pos_emb
+
+ def forward(self, hidden_states: torch.Tensor, grid_spatiotemporal: torch.Tensor) -> torch.Tensor:
+ hidden_states = self.patch_embed(hidden_states)
+ rotary_pos_emb = self.rot_pos_emb(grid_spatiotemporal)
+
+ cu_seqlens = torch.repeat_interleave(
+ grid_spatiotemporal[:, 1] * grid_spatiotemporal[:, 2], grid_spatiotemporal[:, 0]
+ ).cumsum(dim=0, dtype=torch.int32)
+ cu_seqlens = func.pad(cu_seqlens, (1, 0), value=0)
+
+ for blk in self.blocks:
+ hidden_states = blk(hidden_states, cu_seqlens=cu_seqlens, rotary_pos_emb=rotary_pos_emb)
+
+ return self.merger(hidden_states)
+
+
+@add_start_docstrings(
+ "The bare Qwen2VL Model outputting raw hidden-states without any specific head on top.",
+ QWEN2VL_START_DOCSTRING,
+)
+class Qwen2VLModel(Qwen2VLPreTrainedModel):
+ def __init__(self, config: Qwen2VLAConfig):
+ super().__init__(config)
+ self.padding_idx = config.pad_token_id
+ self.vocab_size = config.vocab_size
+
+ self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size, self.padding_idx)
+ self.layers = nn.ModuleList(
+ [Qwen2VLDecoderLayer(config, layer_idx) for layer_idx in range(config.num_hidden_layers)]
+ )
+ self._attn_implementation = config._attn_implementation
+ self.norm = Qwen2RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+ self.rotary_emb = Qwen2VLRotaryEmbedding(config=config)
+
+ self.gradient_checkpointing = False
+ # Initialize weights and apply final processing
+ self.post_init()
+
+ def get_input_embeddings(self):
+ return self.embed_tokens
+
+ def set_input_embeddings(self, value):
+ self.embed_tokens = value
+
+ def forward(
+ self,
+ input_ids: torch.LongTensor = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ position_ids: Optional[torch.LongTensor] = None,
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
+ inputs_embeds: Optional[torch.FloatTensor] = None,
+ use_cache: Optional[bool] = None,
+ output_attentions: Optional[bool] = None,
+ output_hidden_states: Optional[bool] = None,
+ return_dict: Optional[bool] = None,
+ cache_position: Optional[torch.LongTensor] = None,
+ ) -> Union[Tuple, BaseModelOutputWithPast]:
+ output_attentions = (
+ output_attentions if output_attentions is not None else self.config.output_attentions
+ )
+ output_hidden_states = (
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+ )
+ use_cache = use_cache if use_cache is not None else self.config.use_cache
+
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+ if (input_ids is None) ^ (inputs_embeds is not None):
+ raise ValueError("You must specify exactly one of input_ids or inputs_embeds")
+
+ if self.gradient_checkpointing and self.training and use_cache:
+ logger.warning_once(
+ "`use_cache=True` is incompatible with gradient checkpointing. Setting `use_cache=False`..."
+ )
+ use_cache = False
+
+ if inputs_embeds is None:
+ inputs_embeds = self.embed_tokens(input_ids)
+
+ if cache_position is None:
+ past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+ cache_position = torch.arange(
+ past_seen_tokens, past_seen_tokens + inputs_embeds.shape[1], device=inputs_embeds.device
+ )
+
+ # the hard coded `3` is for temporal, height and width.
+ if position_ids is None:
+ position_ids = cache_position.view(1, 1, -1).expand(3, inputs_embeds.shape[0], -1)
+ elif position_ids.dim() == 2:
+ position_ids = position_ids[None, ...].expand(3, position_ids.shape[0], -1)
+
+ causal_mask = self._update_causal_mask(
+ attention_mask, inputs_embeds, cache_position, past_key_values, output_attentions
+ )
+
+ hidden_states = inputs_embeds
+
+ # create position embeddings to be shared across the decoder layers
+ position_embeddings = self.rotary_emb(hidden_states, position_ids)
+
+ # decoder layers
+ all_hidden_states = () if output_hidden_states else None
+ all_self_attns = () if output_attentions else None
+ next_decoder_cache = None
+
+ for decoder_layer in self.layers:
+ if output_hidden_states:
+ all_hidden_states += (hidden_states,)
+
+ if self.gradient_checkpointing and self.training:
+ layer_outputs = self._gradient_checkpointing_func(
+ decoder_layer.__call__,
+ hidden_states,
+ causal_mask,
+ position_ids,
+ past_key_values,
+ output_attentions,
+ use_cache,
+ cache_position,
+ position_embeddings,
+ )
+ else:
+ layer_outputs = decoder_layer(
+ hidden_states,
+ attention_mask=causal_mask,
+ position_ids=position_ids,
+ past_key_value=past_key_values,
+ output_attentions=output_attentions,
+ use_cache=use_cache,
+ cache_position=cache_position,
+ position_embeddings=position_embeddings,
+ )
+
+ hidden_states = layer_outputs[0]
+
+ if use_cache:
+ next_decoder_cache = layer_outputs[2 if output_attentions else 1]
+
+ if output_attentions:
+ all_self_attns += (layer_outputs[1],)
+
+ hidden_states = self.norm(hidden_states)
+
+ # add hidden states from the last decoder layer
+ if output_hidden_states:
+ all_hidden_states += (hidden_states,)
+
+ next_cache = next_decoder_cache if use_cache else None
+
+ if not return_dict:
+ return tuple(
+ v for v in [hidden_states, next_cache, all_hidden_states, all_self_attns] if v is not None
+ )
+ return BaseModelOutputWithPast(
+ last_hidden_state=hidden_states,
+ past_key_values=next_cache,
+ hidden_states=all_hidden_states,
+ attentions=all_self_attns,
+ )
+
+ # Copied from transformers.models.phi3.modeling_phi3.Phi3Model._update_causal_mask
+ def _update_causal_mask(
+ self,
+ attention_mask: torch.Tensor,
+ input_tensor: torch.Tensor,
+ cache_position: torch.Tensor,
+ past_key_values: Cache,
+ output_attentions: bool,
+ ):
+ if self.config._attn_implementation == "flash_attention_2":
+ if attention_mask is not None and 0.0 in attention_mask:
+ return attention_mask
+ return None
+
+ # For SDPA, when possible, we will rely on its `is_causal` argument instead of its `attn_mask` argument, in
+ # order to dispatch on Flash Attention 2. This feature is not compatible with static cache, as SDPA will fail
+ # to infer the attention mask.
+ past_seen_tokens = past_key_values.get_seq_length() if past_key_values is not None else 0
+ using_static_cache = isinstance(past_key_values, StaticCache)
+ using_sliding_window_cache = isinstance(past_key_values, SlidingWindowCache)
+
+ # When output attentions is True, sdpa implementation's forward method calls the eager implementation's forward
+ if (
+ self.config._attn_implementation == "sdpa"
+ and not (using_static_cache or using_sliding_window_cache)
+ and not output_attentions
+ and AttentionMaskConverter._ignore_causal_mask_sdpa(
+ attention_mask,
+ inputs_embeds=input_tensor,
+ past_key_values_length=past_seen_tokens,
+ sliding_window=self.config.sliding_window,
+ is_training=self.training,
+ )
+ ):
+ return None
+
+ dtype, device = input_tensor.dtype, input_tensor.device
+ min_dtype = torch.finfo(dtype).min
+ sequence_length = input_tensor.shape[1]
+ # SlidingWindowCache or StaticCache
+ if using_sliding_window_cache or using_static_cache:
+ target_length = past_key_values.get_max_cache_shape()
+ # DynamicCache or no cache
+ else:
+ target_length = (
+ attention_mask.shape[-1]
+ if isinstance(attention_mask, torch.Tensor)
+ else past_seen_tokens + sequence_length + 1
+ )
+
+ # In case the provided `attention` mask is 2D, we generate a causal mask here (4D).
+ causal_mask = self._prepare_4d_causal_attention_mask_with_cache_position(
+ attention_mask,
+ sequence_length=sequence_length,
+ target_length=target_length,
+ dtype=dtype,
+ device=device,
+ cache_position=cache_position,
+ batch_size=input_tensor.shape[0],
+ config=self.config,
+ past_key_values=past_key_values,
+ )
+
+ if (
+ self.config._attn_implementation == "sdpa"
+ and attention_mask is not None
+ and attention_mask.device.type == "cuda"
+ and not output_attentions
+ ):
+ # Attend to all tokens in fully masked rows in the causal_mask, for example the relevant first rows when
+ # using left padding. This is required by F.scaled_dot_product_attention memory-efficient attention path.
+ # Details: https://github.com/pytorch/pytorch/issues/110213
+ causal_mask = AttentionMaskConverter._unmask_unattended(causal_mask, min_dtype)
+
+ return causal_mask
+
+ @staticmethod
+ # Copied from transformers.models.mistral.modeling_mistral.MistralModel._prepare_4d_causal_attention_mask_with_cache_position with Mistral->Qwen2VL
+ def _prepare_4d_causal_attention_mask_with_cache_position(
+ attention_mask: torch.Tensor,
+ sequence_length: int,
+ target_length: int,
+ dtype: torch.dtype,
+ device: torch.device,
+ cache_position: torch.Tensor,
+ batch_size: int,
+ config: Qwen2VLAConfig,
+ past_key_values: Cache,
+ ):
+ """
+ Creates a causal 4D mask of shape `(batch_size, 1, query_length, key_value_length)` from a 2D mask of shape
+ `(batch_size, key_value_length)`, or if the input `attention_mask` is already 4D, do nothing.
+
+ Args:
+ attention_mask (`torch.Tensor`):
+ A 2D attention mask of shape `(batch_size, key_value_length)` or a 4D attention mask of shape `(batch_size, 1, query_length, key_value_length)`.
+ sequence_length (`int`):
+ The sequence length being processed.
+ target_length (`int`):
+ The target length: when generating with static cache, the mask should be as long as the static cache, to account for the 0 padding, the part of the cache that is not filled yet.
+ dtype (`torch.dtype`):
+ The dtype to use for the 4D attention mask.
+ device (`torch.device`):
+ The device to place the 4D attention mask on.
+ cache_position (`torch.Tensor`):
+ Indices depicting the position of the input sequence tokens in the sequence.
+ batch_size (`torch.Tensor`):
+ Batch size.
+ config (`Qwen2VLConfig`):
+ The model's configuration class
+ past_key_values (`Cache`):
+ The cache class that is being used currently to generate
+ """
+ if attention_mask is not None and attention_mask.dim() == 4:
+ # In this case we assume that the mask comes already in inverted form and requires no inversion or slicing.
+ causal_mask = attention_mask
+ else:
+ min_dtype = torch.finfo(dtype).min
+ causal_mask = torch.full(
+ (sequence_length, target_length), fill_value=min_dtype, dtype=dtype, device=device
+ )
+ diagonal_attend_mask = torch.arange(target_length, device=device) > cache_position.reshape(-1, 1)
+ if config.sliding_window is not None and (
+ not isinstance(past_key_values, SlidingWindowCache) or sequence_length > target_length
+ ):
+ # if we have sliding window, we should not attend to tokens beyond sliding window length, so we mask them out also
+ # the check is needed to verify is current checkpoint was trained with sliding window or not
+ sliding_attend_mask = torch.arange(target_length, device=device) <= (
+ cache_position.reshape(-1, 1) - config.sliding_window
+ )
+ diagonal_attend_mask |= sliding_attend_mask
+ causal_mask *= diagonal_attend_mask
+ causal_mask = causal_mask[None, None, :, :].expand(batch_size, 1, -1, -1)
+ if attention_mask is not None:
+ causal_mask = causal_mask.clone() # copy to contiguous memory for in-place edit
+ if attention_mask.shape[-1] > target_length:
+ attention_mask = attention_mask[:, :target_length]
+ mask_length = attention_mask.shape[-1]
+ padding_mask = causal_mask[:, :, :, :mask_length] + attention_mask[:, None, None, :]
+ padding_mask = padding_mask == 0
+ causal_mask[:, :, :, :mask_length] = causal_mask[:, :, :, :mask_length].masked_fill(
+ padding_mask, min_dtype
+ )
+ return causal_mask
+
+
+QWEN2_VL_INPUTS_DOCSTRING = r"""
+ Args:
+ input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+ Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+ it.
+
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+ [`PreTrainedTokenizer.__call__`] for details.
+
+ [What are input IDs?](../glossary#input-ids)
+ attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+ - 1 for tokens that are **not masked**,
+ - 0 for tokens that are **masked**.
+
+ [What are attention masks?](../glossary#attention-mask)
+
+ Indices can be obtained using [`AutoTokenizer`]. See [`PreTrainedTokenizer.encode`] and
+ [`PreTrainedTokenizer.__call__`] for details.
+
+ If `past_key_values` is used, optionally only the last `decoder_input_ids` have to be input (see
+ `past_key_values`).
+
+ If you want to change padding behavior, you should read [`modeling_opt._prepare_decoder_attention_mask`]
+ and modify to your needs. See diagram 1 in [the paper](https://arxiv.org/abs/1910.13461) for more
+ information on the default strategy.
+
+ - 1 indicates the head is **not masked**,
+ - 0 indicates the head is **masked**.
+ position_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+ Indices of positions of each input sequence tokens in the position embeddings. Selected in the range `[0,
+ config.n_positions - 1]`. [What are position IDs?](../glossary#position-ids)
+ past_key_values (`tuple(tuple(torch.FloatTensor))`, *optional*, returned when `use_cache=True` is passed or when `config.use_cache=True`):
+ Tuple of `tuple(torch.FloatTensor)` of length `config.n_layers`, with each tuple having 2 tensors of shape
+ `(batch_size, num_heads, sequence_length, embed_size_per_head)`) and 2 additional tensors of shape
+ `(batch_size, num_heads, encoder_sequence_length, embed_size_per_head)`.
+
+ Contains pre-computed hidden-states (key and values in the self-attention blocks and in the cross-attention
+ blocks) that can be used (see `past_key_values` input) to speed up sequential decoding.
+
+ If `past_key_values` are used, the user can optionally input only the last `decoder_input_ids` (those that
+ don't have their past key value states given to this model) of shape `(batch_size, 1)` instead of all
+ `decoder_input_ids` of shape `(batch_size, sequence_length)`.
+ inputs_embeds (`torch.FloatTensor` of shape `(batch_size, sequence_length, hidden_size)`, *optional*):
+ Optionally, instead of passing `input_ids` you can choose to directly pass an embedded representation. This
+ is useful if you want more control over how to convert `input_ids` indices into associated vectors than the
+ model's internal embedding lookup matrix.
+ use_cache (`bool`, *optional*):
+ If set to `True`, `past_key_values` key value states are returned and can be used to speed up decoding (see
+ `past_key_values`).
+ output_attentions (`bool`, *optional*):
+ Whether or not to return the attentions tensors of all attention layers. See `attentions` under returned
+ tensors for more detail.
+ output_hidden_states (`bool`, *optional*):
+ Whether or not to return the hidden states of all layers. See `hidden_states` under returned tensors for
+ more detail.
+ return_dict (`bool`, *optional*):
+ Whether or not to return a [`~policy_heads.ModelOutput`] instead of a plain tuple.
+ pixel_values (`torch.FloatTensor` of shape `(seq_length, num_channels * image_size * image_size)):
+ The tensors corresponding to the input images. Pixel values can be obtained using
+ [`AutoImageProcessor`]. See [`Qwen2VLImageProcessor.__call__`] for details. [`Qwen2VLProcessor`] uses
+ [`Qwen2VLImageProcessor`] for processing images.
+ pixel_values_videos (`torch.FloatTensor` of shape `(seq_length, num_channels * temporal_size * image_size * image_size)):
+ The tensors corresponding to the input videos. Pixel values can be obtained using
+ [`AutoImageProcessor`]. See [`Qwen2VLImageProcessor.__call__`] for details. [`Qwen2VLProcessor`] uses
+ [`Qwen2VLImageProcessor`] for processing videos.
+ image_grid_spatiotemporal (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+ The temporal, height and width of feature shape of each image in LLM.
+ video_grid_spatiotemporal (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+ The temporal, height and width of feature shape of each video in LLM.
+ rope_deltas (`torch.LongTensor` of shape `(batch_size, )`, *optional*):
+ The rope index difference between sequence length and multimodal rope.
+"""
+
+
+class Qwen2VLForConditionalGenerationForVLA(Qwen2VLPreTrainedModel, GenerationMixin):
+ _tied_weights_keys = ["lm_head.weight"]
+
+ def __init__(self, config):
+ super().__init__(config)
+ self.visual = Qwen2VisionTransformerPretrainedModel._from_config(
+ config.vision_config, attn_implementation=config._attn_implementation
+ )
+ self.model = Qwen2VLModel(config)
+ self.vocab_size = config.vocab_size
+ self.with_llm_head = config.with_llm_head
+
+ self.padding_side = "left" # set it to left by default, user can use setter to change padding_sides
+ self.using_film = config.using_film
+
+ self.llm_loss_weight = config.llm_loss_weight
+
+ if isinstance(config.policy_head_config, dict):
+ config.policy_head_config = AutoConfig.for_model(**config.policy_head_config)
+ self.policy_head = AutoModel.from_config(config=config.policy_head_config)
+
+ self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+ # Initialize weights and apply final processing
+ self.post_init()
+ if config.policy_head_config.model_type == "scale_dp_policy":
+ self.policy_head.init_weights()
+ self.input_action_proj = ActionProjector(config.hidden_size, config.hidden_size)
+
+ if self.using_film:
+ # Initialize projection layers and condition modulation layers
+ self.reasoning_action_proj = ActionProjector(config.hidden_size, config.hidden_size)
+ self.reasoning_film = FiLM(feature_dim=config.hidden_size, condition_dim=config.hidden_size)
+
+ def get_input_embeddings(self):
+ return self.model.embed_tokens
+
+ def set_input_embeddings(self, value):
+ self.model.embed_tokens = value
+
+ def get_output_embeddings(self):
+ return self.lm_head
+
+ def set_output_embeddings(self, new_embeddings):
+ self.lm_head = new_embeddings
+
+ def set_decoder(self, decoder):
+ self.model = decoder
+
+ def get_decoder(self):
+ return self.model
+
+ def get_rope_index(
+ self,
+ input_ids: torch.LongTensor,
+ image_grid_spatiotemporal: Optional[torch.LongTensor] = None,
+ video_grid_spatiotemporal: Optional[torch.LongTensor] = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ ) -> Tuple[torch.Tensor, torch.Tensor]:
+ """
+ Calculate the 3D rope index based on image and video's temporal, height and width in LLM.
+
+ Explanation:
+ Each embedding sequence contains vision embedding and text embedding or just contains text embedding.
+
+ For pure text embedding sequence, the rotary position embedding has no difference with modern LLMs.
+ Examples:
+ input_ids: [T T T T T], here T is for text.
+ temporal position_ids: [0, 1, 2, 3, 4]
+ height position_ids: [0, 1, 2, 3, 4]
+ width position_ids: [0, 1, 2, 3, 4]
+
+ For vision and text embedding sequence, we calculate 3D rotary position embedding for vision part
+ and 1D rotary position embeddin for text part.
+ Examples:
+ Assume we have a video input with 3 temporal patches, 2 height patches and 2 width patches.
+ input_ids: [V V V V V V V V V V V V T T T T T], here V is for vision.
+ vision temporal position_ids: [0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2]
+ vision height position_ids: [0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1]
+ vision width position_ids: [0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1]
+ text temporal position_ids: [3, 4, 5, 6, 7]
+ text height position_ids: [3, 4, 5, 6, 7]
+ text width position_ids: [3, 4, 5, 6, 7]
+ Here we calculate the text start position_ids as the max vision position_ids plus 1.
+
+ Args:
+ input_ids (`torch.LongTensor` of shape `(batch_size, sequence_length)`):
+ Indices of input sequence tokens in the vocabulary. Padding will be ignored by default should you provide
+ it.
+ image_grid_spatiotemporal (`torch.LongTensor` of shape `(num_images, 3)`, *optional*):
+ The temporal, height and width of feature shape of each image in LLM.
+ video_grid_spatiotemporal (`torch.LongTensor` of shape `(num_videos, 3)`, *optional*):
+ The temporal, height and width of feature shape of each video in LLM.
+ attention_mask (`torch.Tensor` of shape `(batch_size, sequence_length)`, *optional*):
+ Mask to avoid performing attention on padding token indices. Mask values selected in `[0, 1]`:
+
+ - 1 for tokens that are **not masked**,
+ - 0 for tokens that are **masked**.
+
+ Returns:
+ position_ids (`torch.LongTensor` of shape `(3, batch_size, sequence_length)`)
+ mrope_position_deltas (`torch.Tensor` of shape `(batch_size)`)
+ """
+ spatial_merge_size = self.config.vision_config.spatial_merge_size
+ image_token_id = self.config.image_token_id
+ video_token_id = self.config.video_token_id
+ vision_start_token_id = self.config.vision_start_token_id
+ mrope_position_deltas = []
+ if image_grid_spatiotemporal is not None or video_grid_spatiotemporal is not None:
+ total_input_ids = input_ids
+ position_ids = torch.ones(
+ 3, input_ids.shape[0], input_ids.shape[1], dtype=input_ids.dtype, device=input_ids.device
+ )
+ image_index, video_index = 0, 0
+ for i, input_ids in enumerate(total_input_ids):
+ if attention_mask is not None:
+ input_ids = input_ids[attention_mask[i] == 1]
+ image_nums, video_nums = 0, 0
+ vision_start_indices = torch.argwhere(input_ids == vision_start_token_id).squeeze(1)
+ vision_tokens = input_ids[vision_start_indices + 1]
+ image_nums = (vision_tokens == image_token_id).sum()
+ video_nums = (vision_tokens == video_token_id).sum()
+ input_tokens = input_ids.tolist()
+ llm_pos_ids_list: list = []
+ st = 0
+ remain_images, remain_videos = image_nums, video_nums
+ for _ in range(image_nums + video_nums):
+ if image_token_id in input_tokens and remain_images > 0:
+ ed_image = input_tokens.index(image_token_id, st)
+ else:
+ ed_image = len(input_tokens) + 1
+ if video_token_id in input_tokens and remain_videos > 0:
+ ed_video = input_tokens.index(video_token_id, st)
+ else:
+ ed_video = len(input_tokens) + 1
+ if ed_image < ed_video:
+ t, h, w = (
+ image_grid_spatiotemporal[image_index][0],
+ image_grid_spatiotemporal[image_index][1],
+ image_grid_spatiotemporal[image_index][2],
+ )
+ image_index += 1
+ remain_images -= 1
+ ed = ed_image
+ else:
+ t, h, w = (
+ video_grid_spatiotemporal[video_index][0],
+ video_grid_spatiotemporal[video_index][1],
+ video_grid_spatiotemporal[video_index][2],
+ )
+ video_index += 1
+ remain_videos -= 1
+ ed = ed_video
+ llm_grid_t, llm_grid_h, llm_grid_w = (
+ t.item(),
+ h.item() // spatial_merge_size,
+ w.item() // spatial_merge_size,
+ )
+ text_len = ed - st
+
+ st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
+ llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)
+
+ t_index = (
+ torch.arange(llm_grid_t).view(-1, 1).expand(-1, llm_grid_h * llm_grid_w).flatten()
+ )
+ h_index = (
+ torch.arange(llm_grid_h).view(1, -1, 1).expand(llm_grid_t, -1, llm_grid_w).flatten()
+ )
+ w_index = (
+ torch.arange(llm_grid_w).view(1, 1, -1).expand(llm_grid_t, llm_grid_h, -1).flatten()
+ )
+ llm_pos_ids_list.append(torch.stack([t_index, h_index, w_index]) + text_len + st_idx)
+ st = ed + llm_grid_t * llm_grid_h * llm_grid_w
+
+ if st < len(input_tokens):
+ st_idx = llm_pos_ids_list[-1].max() + 1 if len(llm_pos_ids_list) > 0 else 0
+ text_len = len(input_tokens) - st
+ llm_pos_ids_list.append(torch.arange(text_len).view(1, -1).expand(3, -1) + st_idx)
+
+ llm_positions = torch.cat(llm_pos_ids_list, dim=1).reshape(3, -1)
+ position_ids[..., i, attention_mask[i] == 1] = llm_positions.to(position_ids.device)
+ mrope_position_deltas.append(llm_positions.max() + 1 - len(total_input_ids[i]))
+ mrope_position_deltas = torch.tensor(mrope_position_deltas, device=input_ids.device).unsqueeze(1)
+ return position_ids, mrope_position_deltas
+ else:
+ if attention_mask is not None:
+ position_ids = attention_mask.long().cumsum(-1) - 1
+ position_ids.masked_fill_(attention_mask == 0, 1)
+ position_ids = position_ids.unsqueeze(0).expand(3, -1, -1).to(input_ids.device)
+ max_position_ids = position_ids.max(0, keepdim=False)[0].max(-1, keepdim=True)[0]
+ mrope_position_deltas = max_position_ids + 1 - attention_mask.shape[-1]
+ else:
+ position_ids = (
+ torch.arange(input_ids.shape[1], device=input_ids.device)
+ .view(1, 1, -1)
+ .expand(3, input_ids.shape[0], -1)
+ )
+ mrope_position_deltas = torch.zeros(
+ [input_ids.shape[0], 1],
+ device=input_ids.device,
+ dtype=input_ids.dtype,
+ )
+
+ return position_ids, mrope_position_deltas
+
+ def _update_model_kwargs_for_generation(
+ self,
+ outputs: ModelOutput,
+ model_kwargs: Dict[str, Any],
+ is_encoder_decoder: bool = False,
+ num_new_tokens: int = 1,
+ ) -> Dict[str, Any]:
+ model_kwargs = super()._update_model_kwargs_for_generation(
+ outputs=outputs,
+ model_kwargs=model_kwargs,
+ is_encoder_decoder=is_encoder_decoder,
+ num_new_tokens=num_new_tokens,
+ )
+
+ if getattr(outputs, "rope_deltas", None) is not None:
+ model_kwargs["rope_deltas"] = outputs.rope_deltas
+
+ return model_kwargs
+
+ @add_start_docstrings_to_model_forward(QWEN2_VL_INPUTS_DOCSTRING)
+ @replace_return_docstrings(output_type=Qwen2VLCausalLMOutputWithPast, config_class=_CONFIG_FOR_DOC)
+ def forward(
+ self,
+ input_ids: torch.LongTensor = None,
+ attention_mask: Optional[torch.Tensor] = None,
+ position_ids: Optional[torch.LongTensor] = None,
+ past_key_values: Optional[List[torch.FloatTensor]] = None,
+ inputs_embeds: Optional[torch.FloatTensor] = None,
+ labels: Optional[torch.LongTensor] = None,
+ use_cache: Optional[bool] = None,
+ output_attentions: Optional[bool] = None,
+ output_hidden_states: Optional[bool] = None,
+ return_dict: Optional[bool] = None,
+ pixel_values: Optional[torch.Tensor] = None,
+ pixel_values_videos: Optional[torch.FloatTensor] = None,
+ image_grid_spatiotemporal: Optional[torch.LongTensor] = None,
+ video_grid_spatiotemporal: Optional[torch.LongTensor] = None,
+ rope_deltas: Optional[torch.LongTensor] = None,
+ actions: Optional[torch.LongTensor] = None,
+ states: Optional[torch.FloatTensor] = None,
+ is_pad: bool = False,
+ is_eval: bool = False,
+ tinyvla: bool = False,
+ ) -> Union[Tuple, Qwen2VLCausalLMOutputWithPast]:
+ r"""
+ Args:
+ labels (`torch.LongTensor` of shape `(batch_size, sequence_length)`, *optional*):
+ Labels for computing the masked language modeling loss. Indices should either be in `[0, ...,
+ config.vocab_size]` or -100 (see `input_ids` docstring). Tokens with indices set to `-100` are ignored
+ (masked), the loss is only computed for the tokens with labels in `[0, ..., config.vocab_size]`.
+
+ Returns:
+
+ Example:
+
+ ```python
+ >>> from PIL import Image
+ >>> import requests
+ >>> from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
+
+ >>> model = Qwen2VLForConditionalGeneration.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
+ >>> processor = AutoProcessor.from_pretrained("Qwen/Qwen2-VL-7B-Instruct")
+
+ >>> messages = [
+ {
+ "role": "user",
+ "content": [
+ {"type": "image"},
+ {"type": "text", "text": "What is shown in this image?"},
+ ],
+ },
+ ]
+ >>> url = "https://www.ilankelman.org/stopsigns/australia.jpg"
+ >>> image = Image.open(requests.get(url, stream=True).raw)
+
+ >>> text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+ >>> inputs = processor(text=[text], images=[image], vision_infos=[vision_infos])
+
+ >>> # Generate
+ >>> generate_ids = model.generate(inputs.input_ids, max_length=30)
+ >>> tokenizer.batch_decode(generate_ids, skip_special_tokens=True, clean_up_tokenization_spaces=False)[0]
+ "The image shows a street scene with a red stop sign in the foreground. In the background, there is a large red gate with Chinese characters ..."
+ ```"""
+
+ self.computed_type = torch.bfloat16
+ input_ids = input_ids.to("cuda")
+ attention_mask = attention_mask.to("cuda")
+ if not is_eval:
+ labels = labels.to("cuda")
+ actions = actions.to(dtype=self.computed_type, device="cuda")
+ states = states.to(dtype=self.computed_type, device="cuda")
+ position_ids, rope_deltas = self.get_rope_index(
+ input_ids, image_grid_spatiotemporal, video_grid_spatiotemporal, attention_mask
+ )
+ if pixel_values is not None:
+ pixel_values = pixel_values.to(dtype=self.computed_type, device="cuda")
+ output_attentions = (
+ output_attentions if output_attentions is not None else self.config.output_attentions
+ )
+ output_hidden_states = (
+ output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+ )
+ return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+ if inputs_embeds is None:
+ inputs_embeds = self.model.embed_tokens(input_ids)
+ if pixel_values is not None:
+ pixel_values = pixel_values.type(self.visual.get_dtype())
+ image_embeds = self.visual(pixel_values, grid_spatiotemporal=image_grid_spatiotemporal)
+ n_image_tokens = (input_ids == self.config.image_token_id).sum().item()
+ n_image_features = image_embeds.shape[0]
+ if n_image_tokens != n_image_features:
+ raise ValueError(
+ f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+ )
+ image_mask = (
+ (input_ids == self.config.image_token_id)
+ .unsqueeze(-1)
+ .expand_as(inputs_embeds)
+ .to(inputs_embeds.device)
+ )
+ image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+ inputs_embeds = inputs_embeds.masked_scatter(image_mask, image_embeds)
+
+ if pixel_values_videos is not None:
+ pixel_values_videos = pixel_values_videos.type(self.visual.get_dtype())
+ video_embeds = self.visual(pixel_values_videos, grid_spatiotemporal=video_grid_spatiotemporal)
+ n_video_tokens = (input_ids == self.config.video_token_id).sum().item()
+ n_video_features = video_embeds.shape[0]
+ if n_video_tokens != n_video_features:
+ raise ValueError(
+ f"Video features and video tokens do not match: tokens: {n_video_tokens}, features {n_video_features}"
+ )
+ video_mask = (
+ (input_ids == self.config.video_token_id)
+ .unsqueeze(-1)
+ .expand_as(inputs_embeds)
+ .to(inputs_embeds.device)
+ )
+ video_embeds = video_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+ inputs_embeds = inputs_embeds.masked_scatter(video_mask, video_embeds)
+
+ if attention_mask is not None:
+ attention_mask = attention_mask.to(inputs_embeds.device)
+
+ outputs = self.model(
+ input_ids=None,
+ position_ids=position_ids,
+ attention_mask=attention_mask,
+ past_key_values=past_key_values,
+ inputs_embeds=inputs_embeds,
+ use_cache=use_cache,
+ output_attentions=output_attentions,
+ output_hidden_states=True,
+ return_dict=return_dict,
+ )
+
+ hidden_states = outputs[0]
+ if tinyvla: # dex-vla supports tinyvla-style VLA
+ return hidden_states
+
+ if self.with_llm_head:
+ logits = self.lm_head(hidden_states)
+ logits = logits.float()
+ else:
+ logits = None
+ self.llm_head = None
+
+ llm_loss = None
+ # cross-entropy loss for VLM
+ if labels is not None and self.with_llm_head:
+ # Shift so that tokens < n predict n
+ shift_logits = logits[..., :-1, :].contiguous()
+ shift_labels = labels[..., 1:].contiguous()
+ # Flatten the tokens
+ loss_fct = CrossEntropyLoss()
+ shift_logits = shift_logits.view(-1, self.config.vocab_size)
+ shift_labels = shift_labels.view(-1)
+ # Enable model parallelism
+ shift_labels = shift_labels.to(shift_logits.device)
+ llm_loss = loss_fct(shift_logits, shift_labels)
+
+ # for evaluation
+ if is_eval:
+ loss = None
+ if not return_dict:
+ output = (logits,) + outputs[1:]
+ return (loss,) + output if loss is not None else output
+
+ return Qwen2VLCausalLMOutputWithPast(
+ loss=loss,
+ logits=logits,
+ past_key_values=outputs.past_key_values,
+ hidden_states=outputs.hidden_states,
+ attentions=outputs.attentions,
+ rope_deltas=rope_deltas,
+ )
+
+ if self.using_film:
+ action_hidden_states = self.film_forward(
+ labels=labels, input_ids=input_ids, hidden_states=hidden_states
+ )
+ else: # tinyvla
+ action_hidden_states = hidden_states
+
+ ret = self.policy_head(
+ actions=actions, hidden_states=action_hidden_states, states=states, is_pad=is_pad
+ )
+
+ if self.with_llm_head:
+ loss = {
+ "loss": ret["loss"] + self.llm_loss_weight * llm_loss,
+ "llm_loss": llm_loss,
+ "action_loss": ret["loss"],
+ }
+ else:
+ loss = {
+ "loss": ret["loss"],
+ "llm_loss": (torch.ones(1) * (-100)).to(ret["loss"].dtype).squeeze(0),
+ "action_loss": ret["loss"],
+ }
+
+ if not return_dict:
+ output = (logits,) + outputs[1:]
+ return (loss,) + output if loss is not None else output
+
+ torch.cuda.empty_cache()
+ gc.collect()
+ del input_ids
+ del attention_mask
+ del position_ids
+ del past_key_values
+ del inputs_embeds
+ del labels
+ del pixel_values
+ del image_grid_spatiotemporal
+ del actions
+ del states
+ return Qwen2VLCausalLMOutputWithPast(
+ loss=loss,
+ logits=logits,
+ past_key_values=outputs.past_key_values,
+ hidden_states=outputs.hidden_states,
+ attentions=outputs.attentions,
+ rope_deltas=rope_deltas,
+ )
+
+ def film_forward(self, labels, input_ids, hidden_states):
+ """
+ Perform the forward pass for the film module.
+ """
+ inputs_index = labels[:, :] == -100
+ inputs_index = inputs_index.int()
+
+ xor_array = torch.bitwise_xor(inputs_index[:, :-1], inputs_index[:, 1:])
+ indexes = torch.argmax((xor_array != 0).float(), dim=1)
+ input_embeddings = []
+ reasoning_embeddings = []
+ identity = []
+ for i in range(indexes.shape[0]):
+ end = indexes[i] + 1
+ temp = input_ids[i] == 151643 # pad token id for qwen2_vl
+ start = sum(temp.int())
+ input_embeddings.append(self.input_action_proj(hidden_states[i, start:end, :]))
+ identity.append(torch.mean(hidden_states[i, start:end, :], dim=0))
+
+ reasoning_embeddings.append(self.reasoning_action_proj(hidden_states[i, end:, :]))
+ input_embeddings = torch.cat(input_embeddings, dim=0)
+ reasoning_embeddings = torch.cat(reasoning_embeddings, dim=0)
+ identity = torch.stack(identity)
+
+ action_hidden_states = self.reasoning_film(input_embeddings, reasoning_embeddings).unsqueeze(1)
+
+ action_hidden_states = action_hidden_states + identity.unsqueeze(1)
+ return action_hidden_states
+
+ def prepare_inputs_for_generation(
+ self,
+ input_ids,
+ past_key_values=None,
+ attention_mask=None,
+ inputs_embeds=None,
+ cache_position=None,
+ position_ids=None,
+ use_cache=True,
+ pixel_values=None,
+ pixel_values_videos=None,
+ image_grid_spatiotemporal=None,
+ video_grid_spatiotemporal=None,
+ **kwargs,
+ ):
+ # If we have cache: let's slice `input_ids` through `cache_position`, to keep only the unprocessed tokens
+ # Exception 1: when passing input_embeds, input_ids may be missing entries
+ # Exception 2: some generation methods do special slicing of input_ids, so we don't need to do it here
+ if past_key_values is not None:
+ if inputs_embeds is not None: # Exception 1
+ input_ids = input_ids[:, -cache_position.shape[0] :]
+ elif (
+ input_ids.shape[1] != cache_position.shape[0]
+ ): # Default case (the "else", a no op, is Exception 2)
+ input_ids = input_ids[:, cache_position]
+
+ rope_deltas = kwargs.get("rope_deltas")
+ if attention_mask is not None and position_ids is None:
+ if cache_position is None or (cache_position is not None and cache_position[0] == 0):
+ position_ids, rope_deltas = self.get_rope_index(
+ input_ids, image_grid_spatiotemporal, video_grid_spatiotemporal, attention_mask
+ )
+ else:
+ batch_size, seq_length = input_ids.shape
+ delta = (
+ cache_position[0] + rope_deltas
+ if cache_position is not None and rope_deltas is not None
+ else 0
+ )
+ position_ids = torch.arange(seq_length, device=input_ids.device)
+ position_ids = position_ids.view(1, -1).expand(batch_size, -1)
+ position_ids = position_ids.add(delta)
+ position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)
+
+ if cache_position[0] != 0:
+ pixel_values = None
+ pixel_values_videos = None
+
+ # if `inputs_embeds` are passed, we only want to use them in the 1st generation step
+ if inputs_embeds is not None and cache_position[0] == 0:
+ model_inputs = {"inputs_embeds": inputs_embeds, "input_ids": None}
+ else:
+ model_inputs = {"input_ids": input_ids, "inputs_embeds": None}
+
+ if isinstance(past_key_values, StaticCache) and attention_mask.ndim == 2:
+ if model_inputs["inputs_embeds"] is not None:
+ batch_size, sequence_length, _ = inputs_embeds.shape
+ device = inputs_embeds.device
+ else:
+ batch_size, sequence_length = input_ids.shape
+ device = input_ids.device
+
+ attention_mask = self.model._prepare_4d_causal_attention_mask_with_cache_position(
+ attention_mask,
+ sequence_length=sequence_length,
+ target_length=past_key_values.get_max_cache_shape(),
+ dtype=self.lm_head.weight.dtype,
+ device=device,
+ cache_position=cache_position,
+ batch_size=batch_size,
+ config=self.config,
+ past_key_values=past_key_values,
+ )
+
+ model_inputs.update(
+ {
+ "position_ids": position_ids,
+ "past_key_values": past_key_values,
+ "use_cache": use_cache,
+ "attention_mask": attention_mask,
+ "pixel_values": pixel_values,
+ "pixel_values_videos": pixel_values_videos,
+ "image_grid_spatiotemporal": image_grid_spatiotemporal,
+ "video_grid_spatiotemporal": video_grid_spatiotemporal,
+ "rope_deltas": rope_deltas,
+ }
+ )
+ model_inputs.update(kwargs)
+ return model_inputs
diff --git a/lerobot/common/policies/dexvla/robot_data_processor.py b/lerobot/common/policies/dexvla/robot_data_processor.py
new file mode 100644
index 00000000..7af0aa05
--- /dev/null
+++ b/lerobot/common/policies/dexvla/robot_data_processor.py
@@ -0,0 +1,172 @@
+#!/usr/bin/env python
+
+# Copyright 2025 DexVLA Team 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.
+
+import numpy as np
+import torch
+from PIL import Image
+from qwen_vl_utils import fetch_image
+
+
+class Qwen2VLAProcess:
+ def __init__(
+ self,
+ tokenizer=None,
+ max_seq_len=512,
+ multimodal_processor=None,
+ ):
+ super().__init__()
+ self.tokenizer = tokenizer
+ self.max_seq_len = max_seq_len
+ self.multimodal_processor = multimodal_processor
+
+ def qwen2_image_preprocess(self, each):
+ ele = {}
+ each = Image.fromarray(each.squeeze(0).permute(1, 2, 0).cpu().numpy().astype(np.uint8))
+ ele["image"] = each
+
+ ele["resized_height"] = each.height
+ ele["resized_width"] = each.width
+ each = fetch_image(ele)
+ return torch.from_numpy(np.array(each))
+
+ def single_forward_process(self, images, raw_lang, reasoning, eval=False, use_reasoning=True):
+ len_views = images.shape[0]
+ messages = self.construct_chat_data(len_views, raw_lang)
+
+ data_dict = {"messages": messages}
+
+ image_data = torch.chunk(images, len_views, 0)
+
+ images_list = []
+
+ for _i, each in enumerate(image_data):
+ img_pil = self.qwen2_image_preprocess(each)
+ images_list.append(img_pil)
+
+ text = self.multimodal_processor.apply_chat_template(
+ messages, tokenize=False, add_generation_prompt=True
+ )
+
+ model_inputs = self.multimodal_processor(
+ text=text,
+ images=images_list,
+ videos=None,
+ padding=True,
+ return_tensors="pt",
+ )
+
+ if eval:
+ new_dict = {}
+ for k, v in model_inputs.items():
+ if "image_grid" in k:
+ new_dict["image_grid_spatiotemporal"] = v
+ else:
+ new_dict[k] = v
+ return new_dict
+
+ input_labels = torch.ones_like(model_inputs["input_ids"]) * -100
+ answer = reasoning + " Next action:" + "<|im_end|>" if use_reasoning else "" + "<|im_end|>"
+
+ output_text = self.tokenizer(answer, padding=True, return_tensors="pt")
+ output_labels = output_text["input_ids"]
+ model_inputs["input_ids"] = torch.cat((model_inputs["input_ids"], output_text["input_ids"]), dim=-1)
+ model_inputs["attention_mask"] = torch.cat(
+ (model_inputs["attention_mask"], output_text["attention_mask"]), dim=-1
+ )
+ labels = torch.cat((input_labels, output_labels), dim=-1)
+
+ data_dict["labels"] = labels
+ for k, v in model_inputs.items():
+ if "image_grid" in k:
+ data_dict["image_grid_spatiotemporal"] = v
+ else:
+ data_dict[k] = v
+ return data_dict
+
+ def forward(self, batch, use_reasoning=True):
+ """This is the main process function for processing vl data into Qwen2_vl format"""
+ all_images = batch["images"]
+ all_images = torch.einsum(
+ "v b c h w -> b v c h w", all_images
+ ) # camera_views, batch_size, channel, height, width
+
+ ret_l = []
+
+ for idx, images in enumerate(all_images):
+ raw_lang = batch["raw_langs"][idx]
+ reasoning = batch["reasonings"][idx]
+ ret_dict = self.single_forward_process(images, raw_lang, reasoning, use_reasoning=use_reasoning)
+ ret_l.append(ret_dict)
+
+ return self.post_process(ret_l)
+
+ def post_process(self, instances):
+ input_ids = [torch.flip(instance["input_ids"].squeeze(0), dims=[0]) for instance in instances]
+ labels = [torch.flip(instance["labels"].squeeze(0), dims=[0]) for instance in instances]
+
+ image_grid_spatiotemporal = torch.stack(
+ [instances["image_grid_spatiotemporal"] for instances in instances]
+ )
+ pixel_values = torch.stack([instances["pixel_values"] for instances in instances])
+ pixel_values_videos = None
+ video_grid_spatiotemporal = None
+
+ labels = torch.nn.utils.rnn.pad_sequence(labels, batch_first=True, padding_value=-100)
+ labels = torch.flip(labels, dims=[1])
+ input_ids = torch.nn.utils.rnn.pad_sequence(
+ input_ids, batch_first=True, padding_value=self.tokenizer.pad_token_id
+ )
+ input_ids = torch.flip(input_ids, dims=[1])
+ b = input_ids.shape[0]
+
+ image_grid_spatiotemporal = image_grid_spatiotemporal.reshape(
+ b * image_grid_spatiotemporal.shape[1], image_grid_spatiotemporal.shape[2]
+ )
+ pixel_values = pixel_values.reshape(b * pixel_values.shape[1], pixel_values.shape[2])
+
+ attention_mask = (input_ids.ne(self.tokenizer.pad_token_id),)
+
+ batch = {
+ "input_ids": input_ids,
+ "attention_mask": attention_mask[0],
+ "labels": labels,
+ "image_grid_spatiotemporal": image_grid_spatiotemporal,
+ "pixel_values_videos": pixel_values_videos,
+ "video_grid_spatiotemporal": video_grid_spatiotemporal,
+ "pixel_values": pixel_values,
+ }
+
+ return batch
+
+ def construct_chat_data(self, len_image, raw_lang):
+ messages = [
+ {
+ "role": "user",
+ "content": [],
+ },
+ ]
+
+ for _i in range(len_image):
+ messages[0]["content"].append(
+ {
+ "type": "image",
+ "image": None,
+ }
+ )
+ messages[0]["content"].append({"type": "text", "text": ""})
+ messages[0]["content"][-1]["text"] = raw_lang
+
+ return messages
diff --git a/lerobot/common/policies/factory.py b/lerobot/common/policies/factory.py
index 8def95a3..4d9b24e8 100644
--- a/lerobot/common/policies/factory.py
+++ b/lerobot/common/policies/factory.py
@@ -23,6 +23,7 @@ from lerobot.common.datasets.utils import dataset_to_policy_features
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.dexvla.configuration_dexvla import DexVLAConfig
from lerobot.common.policies.diffusion.configuration_diffusion import DiffusionConfig
from lerobot.common.policies.pi0.configuration_pi0 import PI0Config
from lerobot.common.policies.pi0fast.configuration_pi0fast import PI0FASTConfig
@@ -55,6 +56,10 @@ def get_policy_class(name: str) -> PreTrainedPolicy:
from lerobot.common.policies.pi0.modeling_pi0 import PI0Policy
return PI0Policy
+ elif name == "dexvla":
+ from lerobot.common.policies.dexvla.modeling_dexvla import DexVLAPolicy
+
+ return DexVLAPolicy
elif name == "pi0fast":
from lerobot.common.policies.pi0fast.modeling_pi0fast import PI0FASTPolicy
@@ -74,6 +79,8 @@ def make_policy_config(policy_type: str, **kwargs) -> PreTrainedConfig:
return VQBeTConfig(**kwargs)
elif policy_type == "pi0":
return PI0Config(**kwargs)
+ elif policy_type == "dexvla":
+ return DexVLAConfig(**kwargs)
elif policy_type == "pi0fast":
return PI0FASTConfig(**kwargs)
else:
diff --git a/pyproject.toml b/pyproject.toml
index 4b858634..361e0857 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -85,6 +85,7 @@ dynamixel = ["dynamixel-sdk>=3.7.31", "pynput>=1.7.7"]
feetech = ["feetech-servo-sdk>=1.0.0", "pynput>=1.7.7"]
intelrealsense = ["pyrealsense2>=2.55.1.6486 ; sys_platform != 'darwin'"]
pi0 = ["transformers>=4.48.0"]
+dexvla = ["transformers>=4.45.2", "qwen_vl_utils==0.0.10", "timm==0.9.10"]
pusht = ["gym-pusht>=0.1.5 ; python_version < '4.0'"]
stretch = [
"hello-robot-stretch-body>=0.7.27 ; python_version < '4.0' and sys_platform == 'linux'",