optimum-rbln 0.8.2a4__py3-none-any.whl → 0.9.3rc0__py3-none-any.whl

optimum/rbln/transformers/models/qwen2_5_vl/qwen2_5_vl_architecture.py

@@ -3,11 +3,9 @@ from typing import Tuple
 
 import torch
 import torch.nn as nn
+from transformers import PreTrainedModel
 
-from ..decoderonly.decoderonly_architecture import (
-    DecoderOnlyWrapper,
-    apply_rotary_pos_emb,
-)
+from ..decoderonly.decoderonly_architecture import DecoderOnlyWrapper, apply_rotary_pos_emb
 
 
 class Qwen2_5_VisionTransformerWrapper(nn.Module):
@@ -159,15 +157,16 @@ class Qwen2_5_VLVisionWindowAttention(nn.Module):
 class Qwen2_5_VL_LanguageModelWrapper(DecoderOnlyWrapper):
     def prepare_forward_args(self, *args):
         args = list(args)
-        input_ids = None if self.use_inputs_embeds else args.pop(0)
-        inputs_embeds = args.pop(0) if self.use_inputs_embeds else None
+        input_ids = None if self.rbln_config.use_inputs_embeds else args.pop(0)
+        inputs_embeds = args.pop(0) if self.rbln_config.use_inputs_embeds else None
         cache_position = args.pop(0)
         global_block_tables = args.pop(0)
         local_block_tables = None
         position_embeds = args.pop(0)
         query_position = args.pop(0) if self.phase == "prefill" else None
         position_ids = None
-        attention_mask = args.pop(0) if self.use_attention_mask else None
+        lora_int_id = None
+        attention_mask = args.pop(0) if self.rbln_config.use_attention_mask else None
         past_key_values = args
 
         if len(past_key_values) != 2 * self.num_hidden_layers:
@@ -194,6 +193,28 @@ class Qwen2_5_VL_LanguageModelWrapper(DecoderOnlyWrapper):
             query_position,
             attention_mask,
             position_ids,
+            lora_int_id,
             past_key_values,
             position_embeds,
         )
+
+    def convert_to_rbln_class(self, model: PreTrainedModel, max_seq_len: int):
+        new_layers = []
+
+        for layer_idx, layer in enumerate(model.model.language_model.layers):
+            is_sliding = layer_idx in self.rbln_config.sliding_window_layers
+            new_self_attn = self.get_rbln_attn_class()(
+                self.get_attn_layer(layer), self.rbln_config, is_sliding=is_sliding
+            )
+            new_layer = self.get_rbln_layer_class()(layer, new_self_attn)
+            new_layers.append(new_layer)
+
+        new_model = self.get_rbln_model_class()(
+            model.model.language_model,
+            new_layers,
+            self.rbln_config,
+            use_learned_pos_emb=self.__class__._use_learned_pos_emb,
+        )
+
+        new_model = self.get_rbln_causal_lm_class()(model.model, new_model)
+        return new_model

optimum/rbln/transformers/models/qwen2_vl/__init__.py

@@ -0,0 +1,19 @@
+# Copyright 2025 Rebellions Inc. All rights reserved.
+
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at:
+
+#     http://www.apache.org/licenses/LICENSE-2.0
+
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+from .configuration_qwen2_vl import (
+    RBLNQwen2VisionTransformerPretrainedModelConfig,
+    RBLNQwen2VLForConditionalGenerationConfig,
+)
+from .modeling_qwen2_vl import RBLNQwen2VisionTransformerPretrainedModel, RBLNQwen2VLForConditionalGeneration

optimum/rbln/transformers/models/qwen2_vl/configuration_qwen2_vl.py

@@ -0,0 +1,88 @@
+# Copyright 2025 Rebellions Inc. 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 typing import Any, Dict, List, Optional, Union
+
+from ....configuration_utils import RBLNModelConfig
+from ..decoderonly.configuration_decoderonly import RBLNDecoderOnlyModelForCausalLMConfig
+
+
+class RBLNQwen2VLForConditionalGenerationConfig(RBLNDecoderOnlyModelForCausalLMConfig):
+    submodules = ["visual"]
+
+    def __init__(
+        self,
+        use_inputs_embeds: bool = True,
+        visual: Optional[RBLNModelConfig] = None,
+        **kwargs: Dict[str, Any],
+    ):
+        """
+        Args:
+            use_inputs_embeds (bool): Whether or not to use `inputs_embeds` as input. Defaults to `True`.
+            visual (Optional[RBLNModelConfig]): Configuration for the vision encoder component.
+            kwargs: Additional arguments passed to the parent `RBLNDecoderOnlyModelForCausalLMConfig`.
+
+        Raises:
+            ValueError: If `use_inputs_embeds` is False.
+            ValueError: If the visual configuration is provided but contains invalid settings, such as an invalid max_seq_lens (e.g., not a positive integer or insufficient for the expected resolution).
+            ValueError: If visual is None and no default vision configuration can be inferred for the model architecture.
+            ValueError: If any inherited parameters violate constraints defined in the parent class, such as batch_size not being a positive integer, prefill_chunk_size not being divisible by 64, or max_seq_len not meeting requirements for Flash Attention.
+        """
+        super().__init__(use_inputs_embeds=use_inputs_embeds, **kwargs)
+        if not self.use_inputs_embeds:
+            raise ValueError(
+                "RBLNQwen2VLForConditionalGenerationConfig does not allow `use_inputs_embeds` to be set to False, "
+                "as RBLNQwen2VLForConditionalGeneration accepts only `inputs_embeds` as input."
+            )
+        self.visual = visual
+
+
+class RBLNQwen2VisionTransformerPretrainedModelConfig(RBLNModelConfig):
+    def __init__(self, max_seq_lens: Union[int, List[int]] = None, **kwargs: Dict[str, Any]):
+        """
+        Args:
+            max_seq_lens (Optional[Union[int, List[int]]]): Maximum sequence lengths for Vision
+                Transformer attention. Can be an integer or list of integers, each indicating
+                the number of patches in a sequence for an image or video. For example, an image
+                of 224x224 pixels with patch size 14 results in (224/14) * (224/14) = 256 patches,
+                so `max_seq_lens` must be at least 256. RBLN optimization runs inference per image
+                or video frame, so set `max_seq_lens` to match the maximum expected resolution to
+                optimize computation. If not provided, a `ValueError` is raised.
+            kwargs: Additional arguments passed to the parent RBLNModelConfig.
+
+        Raises:
+            ValueError: If batch_size is not a positive integer.
+            ValueError: If `max_seq_lens` (or any value in the list) is not a positive integer.
+            ValueError: If `max_seq_lens` is insufficient for the expected image/video resolution.
+            ValueError: If `batch_size` (inherited from RBLNModelConfig) is not a positive integer.
+
+        Max Seq Lens:
+            Since `Qwen2VLForConditionalGeneration` performs inference on a per-image or per-frame basis,
+            `max_seq_lens` should be set based on the maximum expected resolution of the input images or video frames.
+
+            The value must be greater than or equal to the number of patches generated from the input image.
+            For example, a 224x224 image with a patch size of 14 results in (224 / 14) * (224 / 14) = 256 patches.
+            Therefore, `max_seq_lens` must be at least 256.
+        """
+        super().__init__(**kwargs)
+
+        if max_seq_lens is not None:
+            if isinstance(max_seq_lens, int):
+                max_seq_lens = [max_seq_lens]
+            elif isinstance(max_seq_lens, list):
+                max_seq_lens.sort(reverse=True)
+        else:
+            raise ValueError("'max_seq_lens' must be specified.")
+
+        self.max_seq_lens = max_seq_lens

optimum/rbln/transformers/models/qwen2_vl/modeling_qwen2_vl.py

@@ -0,0 +1,514 @@
+# Copyright 2025 Rebellions Inc. 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 inspect
+from pathlib import Path
+from typing import TYPE_CHECKING, Any, Callable, Optional, Union
+
+import torch
+from transformers import (
+    AutoModelForVision2Seq,
+    PretrainedConfig,
+    PreTrainedModel,
+    Qwen2VLForConditionalGeneration,
+)
+from transformers.modeling_utils import no_init_weights
+from transformers.models.qwen2_vl.modeling_qwen2_vl import (
+    PatchEmbed,
+    Qwen2VisionTransformerPretrainedModel,
+    Qwen2VLModel,
+    Qwen2VLRotaryEmbedding,
+    VisionRotaryEmbedding,
+)
+
+from ....configuration_utils import RBLNCompileConfig
+from ....modeling import RBLNModel
+from ....utils.logging import get_logger
+from ..decoderonly.modeling_decoderonly import RBLNDecoderOnlyModelForCausalLM, RBLNDecoderOnlyOutput
+from .configuration_qwen2_vl import (
+    RBLNQwen2VisionTransformerPretrainedModelConfig,
+    RBLNQwen2VLForConditionalGenerationConfig,
+)
+from .qwen2_vl_architecture import Qwen2VisionTransformerWrapper, Qwen2VL_LanguageModelWrapper
+
+
+logger = get_logger(__name__)
+
+if TYPE_CHECKING:
+    from transformers import (
+        AutoFeatureExtractor,
+        AutoProcessor,
+        AutoTokenizer,
+        PretrainedConfig,
+    )
+
+
+class RBLNQwen2VisionTransformerPretrainedModel(RBLNModel):
+    auto_model_class = None
+
+    def __post_init__(self, **kwargs):
+        self.transformer = self.model[0]
+        self.max_seq_lens = torch.tensor(sorted(self.rbln_config.max_seq_lens, reverse=False))
+        config = self.config
+
+        self.patch_size = config.spatial_patch_size
+        self.spatial_merge_size = config.spatial_merge_size
+        self.spatial_merge_unit = config.spatial_merge_size * config.spatial_merge_size
+        self.rotary_pos_emb = VisionRotaryEmbedding((config.embed_dim // config.num_heads) // 2)
+        with no_init_weights():
+            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,
+            ).eval()
+        artifacts = torch.load(self.model_save_dir / self.subfolder / "torch_artifacts.pth", weights_only=False)
+        self.patch_embed.load_state_dict(artifacts["patch_embed"])
+
+    @classmethod
+    def save_torch_artifacts(
+        cls,
+        model: "Qwen2VLForConditionalGeneration",
+        save_dir_path: Path,
+        subfolder: str,
+        rbln_config: RBLNQwen2VisionTransformerPretrainedModelConfig,
+    ):
+        save_dict = {}
+        save_dict["patch_embed"] = model.patch_embed.state_dict()
+        torch.save(save_dict, save_dir_path / subfolder / "torch_artifacts.pth")
+
+    @classmethod
+    def wrap_model_if_needed(
+        cls, model: "PreTrainedModel", rbln_config: RBLNQwen2VisionTransformerPretrainedModelConfig
+    ):
+        return Qwen2VisionTransformerWrapper(model).eval()
+
+    def __getattr__(self, __name: str) -> Any:
+        def redirect(func):
+            return lambda *pargs, **kwargs: func(self, *pargs, **kwargs)
+
+        val = getattr(Qwen2VisionTransformerPretrainedModel, __name)
+
+        if isinstance(val, Callable) and "self" in set(inspect.signature(val).parameters):
+            return redirect(val)
+        return val
+
+    @classmethod
+    def _update_rbln_config(
+        cls,
+        preprocessors: Union["AutoFeatureExtractor", "AutoProcessor", "AutoTokenizer"],
+        model: Optional["PreTrainedModel"] = None,
+        model_config: "PretrainedConfig" = None,
+        rbln_config: Optional[RBLNQwen2VisionTransformerPretrainedModelConfig] = None,
+    ) -> RBLNQwen2VisionTransformerPretrainedModelConfig:
+        hidden_size = getattr(model_config, "embed_dim")
+        num_heads = getattr(model_config, "num_heads")
+        head_dim = hidden_size // num_heads
+
+        input_infos = []
+        for max_seq_len in rbln_config.max_seq_lens:
+            input_info = [
+                ("hidden_states", [max_seq_len, hidden_size], "float32"),
+                ("full_attn_masks", [1, 1, max_seq_len, max_seq_len], "float32"),
+                (
+                    "cos",
+                    [1, 1, max_seq_len, head_dim],
+                    "float32",
+                ),
+                (
+                    "sin",
+                    [1, 1, max_seq_len, head_dim],
+                    "float32",
+                ),
+            ]
+            input_infos.append(input_info)
+
+        rbln_compile_config = RBLNCompileConfig(input_info=input_infos)
+        rbln_config.set_compile_cfgs([rbln_compile_config])
+
+        return rbln_config
+
+    @staticmethod
+    def _pad_for_full_attn_layers(hidden_state, cos, sin, max_seq_len):
+        if hidden_state.shape[0] < max_seq_len:
+            full_padding_size = max_seq_len - hidden_state.shape[0]
+            full_padding_hidden = torch.zeros(
+                full_padding_size,
+                hidden_state.shape[-1],
+                dtype=hidden_state.dtype,
+            )
+            hidden_state_full_padded = torch.cat([hidden_state, full_padding_hidden], dim=0)
+            full_padding_pos = torch.zeros(
+                full_padding_size,
+                cos.shape[-1],
+                dtype=cos.dtype,
+            )
+            cos_full_padded = torch.cat([cos, full_padding_pos], dim=0)
+            sin_full_padded = torch.cat([sin, full_padding_pos], dim=0)
+        else:
+            hidden_state_full_padded = hidden_state
+            cos_full_padded = cos
+            sin_full_padded = sin
+
+        full_attn_masks = torch.ones(
+            1,
+            1,
+            max_seq_len,
+            max_seq_len,
+            dtype=torch.float32,
+        )
+
+        full_attn_masks[:, :, hidden_state.shape[0] : max_seq_len, :] = 0
+        full_attn_masks[:, :, :, hidden_state.shape[0] : max_seq_len] = 0
+        return hidden_state_full_padded, cos_full_padded, sin_full_padded, full_attn_masks
+
+    def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor) -> torch.Tensor:
+        # Processes a batch of images (or frames) through the vision transformer.
+        # Each image is handled independently for padding and attention mask generation.
+
+        hidden_states = self.patch_embed(hidden_states)
+        rotary_pos_emb = self.rot_pos_emb(grid_thw)
+        emb = torch.cat((rotary_pos_emb, rotary_pos_emb), dim=-1)
+        position_embeddings = (emb.cos(), emb.sin())
+
+        cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum(
+            dim=0,
+            dtype=torch.int32,
+        )
+        cu_seqlens = torch.nn.functional.pad(cu_seqlens, (1, 0), value=0)
+
+        num_images = len(cu_seqlens) - 1
+        output_hidden_states = []
+
+        # Process each image in the sequence
+        for i in range(num_images):
+            image_s, image_e = cu_seqlens[i], cu_seqlens[i + 1]
+
+            # Select the nearest higher max_seq_len from the available compiled models.
+            cu_seq_len = image_e - image_s
+            try:
+                cu_index = torch.searchsorted(self.max_seq_lens, cu_seq_len).item()
+                max_seq_len = self.max_seq_lens[cu_index]
+            except Exception:
+                raise ValueError(
+                    f"Required seq_len({cu_seq_len}) is larger than available max_seq_lens({self.max_seq_lens.tolist()})."
+                )
+
+            # Padding for Full Attention Layers
+            hidden_state_full_padded, cos_full_padded, sin_full_padded, full_attn_masks = (
+                self._pad_for_full_attn_layers(
+                    hidden_states[image_s:image_e],
+                    position_embeddings[0][image_s:image_e],
+                    position_embeddings[1][image_s:image_e],
+                    max_seq_len,
+                )
+            )
+
+            # RBLN run with the compiled model
+            output = self.transformer(
+                hidden_state_full_padded,
+                full_attn_masks,
+                cos_full_padded[None, None, :, :],
+                sin_full_padded[None, None, :, :],
+            )
+            # Depadding
+            depadded_output = output[: cu_seq_len // self.spatial_merge_unit]
+            output_hidden_states.append(depadded_output)
+
+        hidden_states = torch.cat(output_hidden_states)
+        return hidden_states
+
+
+class RBLNQwen2VLForConditionalGeneration(RBLNDecoderOnlyModelForCausalLM):
+    """
+    RBLNQwen2VLForConditionalGeneration is a multi-modal model that integrates vision and language processing capabilities,
+    optimized for RBLN NPUs. It is designed for conditional generation tasks that involve both image and text inputs.
+
+    This model inherits from [`RBLNDecoderOnlyModelForCausalLM`]. Check the superclass documentation for the generic methods the library implements for all its models.
+
+    Important Note:
+        This model includes a Large Language Model (LLM). For optimal performance, it is highly recommended to use
+        tensor parallelism for the language model. This can be achieved by using the `rbln_config` parameter in the
+        `from_pretrained` method. Refer to the `from_pretrained` documentation and the RBLNQwen2VLForConditionalGenerationConfig class for details.
+
+    Examples:
+        ```python
+        from optimum.rbln import RBLNQwen2VLForConditionalGeneration
+
+        model = RBLNQwen2VLForConditionalGeneration.from_pretrained(
+            "Qwen/Qwen2-VL-7B-Instruct",
+            export=True,
+            rbln_config={
+                "visual": {
+                    "max_seq_lens": 6400,
+                    "device": 0,
+                },
+                "tensor_parallel_size": 8,
+                "max_seq_len": 32_768,
+                "device": [0, 1, 2, 3, 4, 5, 6, 7],
+            },
+        )
+
+        model.save_pretrained("compiled-qwen2-vl-7b-instruct")
+        ```
+    """
+
+    auto_model_class = AutoModelForVision2Seq
+    _rbln_submodules = [
+        {"name": "visual"},
+    ]
+    _decoder_wrapper_cls = Qwen2VL_LanguageModelWrapper
+    _use_rotary_emb = False
+
+    def __post_init__(self, **kwargs):
+        super().__post_init__(**kwargs)
+        self.visual = self.rbln_submodules[0]
+        self.mrope_section = self.config.rope_scaling["mrope_section"]
+        self.rotary_emb = Qwen2VLRotaryEmbedding(self.config)
+        self.rope_deltas = torch.zeros(self.rbln_config.batch_size)
+
+    def can_generate(self):
+        return True
+
+    @classmethod
+    def get_pytorch_model(cls, *args, **kwargs):
+        model = super().get_pytorch_model(*args, **kwargs)
+        model.model.lm_head = model.lm_head
+        model.lm_head = None
+        del model.lm_head
+        return model
+
+    @classmethod
+    def get_input_info(
+        cls,
+        batch_size: int,
+        query_length: int,
+        rbln_config: RBLNQwen2VLForConditionalGenerationConfig,
+        model_config: PretrainedConfig,
+    ):
+        input_info = super().get_input_info(batch_size, query_length, rbln_config, model_config)
+        pos_idx = 3
+        input_info.insert(
+            pos_idx,
+            (
+                "position_emb",
+                [2, batch_size, 1, query_length, model_config.hidden_size // model_config.num_attention_heads],
+                "float32",
+            ),
+        )
+
+        return input_info
+
+    def prepare_inputs_for_generation(
+        self,
+        input_ids: torch.LongTensor,
+        generate_idx: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.Tensor] = None,
+        pixel_values=None,
+        pixel_values_videos=None,
+        image_grid_thw=None,
+        video_grid_thw=None,
+        **kwargs,
+    ):
+        model_inputs = super().prepare_inputs_for_generation(
+            input_ids,
+            generate_idx,
+            attention_mask,
+            inputs_embeds,
+            **kwargs,
+        )
+
+        is_prefill_phase = generate_idx is None
+        if is_prefill_phase:
+            model_inputs.update({"input_ids": input_ids})
+
+        model_inputs.update(
+            {
+                "pixel_values": pixel_values,
+                "pixel_values_videos": pixel_values_videos,
+                "image_grid_thw": image_grid_thw,
+                "video_grid_thw": video_grid_thw,
+            }
+        )
+
+        return model_inputs
+
+    def _get_position_embeddings(self, hidden_states, position_ids):
+        cos, sin = self.rotary_emb(hidden_states, position_ids)
+        mrope_section = self.mrope_section * 2
+        cos = torch.cat([m[i % 3] for i, m in enumerate(cos.split(mrope_section, dim=-1))], dim=-1).unsqueeze(1)
+        sin = torch.cat([m[i % 3] for i, m in enumerate(sin.split(mrope_section, dim=-1))], dim=-1).unsqueeze(1)
+        return torch.stack([cos, sin])
+
+    def _preprocess_prefill(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: torch.Tensor = None,
+        pixel_values: torch.Tensor = None,
+        pixel_values_videos: torch.FloatTensor = None,
+        image_grid_thw: torch.LongTensor = None,
+        video_grid_thw: torch.LongTensor = None,
+    ):
+        batch_size = input_ids.shape[0]
+        inputs_embeds = self.embed_tokens(input_ids)
+
+        if pixel_values is not None:
+            image_embeds = self.visual(pixel_values, grid_thw=image_grid_thw)
+            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}"
+                )
+
+            mask = input_ids == self.config.image_token_id
+            mask_unsqueezed = mask.unsqueeze(-1)
+            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
+
+            image_embeds = image_embeds.to(inputs_embeds.device, inputs_embeds.dtype)
+            inputs_embeds = inputs_embeds.masked_scatter(mask_expanded, image_embeds)
+
+        if pixel_values_videos is not None:
+            video_embeds = self.visual(pixel_values_videos, grid_thw=video_grid_thw)
+            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}"
+                )
+
+            mask = input_ids == self.config.video_token_id
+            mask_unsqueezed = mask.unsqueeze(-1)
+            mask_expanded = mask_unsqueezed.expand_as(inputs_embeds)
+            inputs_embeds = inputs_embeds.masked_scatter(mask_expanded, video_embeds)
+
+        max_inputs_len = input_ids.shape[1]
+
+        head_dim = getattr(self.config, "head_dim", None) or self.config.hidden_size // self.config.num_attention_heads
+        all_position_embeds = torch.zeros(2, batch_size, 1, max_inputs_len, head_dim)
+        all_rope_deltas = []
+
+        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
+        image_idx, video_idx = 0, 0
+
+        for b_idx in range(batch_size):
+            input_id = input_ids[b_idx : b_idx + 1][:, attention_mask[b_idx].bool()]
+            vision_start_indices = torch.argwhere(input_id == vision_start_token_id).squeeze(1)
+            vision_tokens = input_id[0][vision_start_indices + 1]
+            image_nums = (vision_tokens == image_token_id).sum()
+            video_nums = (vision_tokens == video_token_id).sum()
+            position_ids, rope_deltas = Qwen2VLModel.get_rope_index(
+                self,
+                input_id,
+                image_grid_thw[image_idx : image_idx + image_nums] if image_grid_thw is not None else None,
+                video_grid_thw[video_idx : video_idx + video_nums] if video_grid_thw is not None else None,
+            )
+            image_idx += image_nums
+            video_idx += video_nums
+
+            position_embed = self._get_position_embeddings(inputs_embeds, position_ids)
+            mask_indices = torch.nonzero(attention_mask[b_idx], as_tuple=True)[0]
+            all_position_embeds[:, b_idx : b_idx + 1].index_copy_(dim=-2, index=mask_indices, source=position_embed)
+            all_rope_deltas.append(rope_deltas)
+
+        rope_deltas = torch.stack(all_rope_deltas)
+
+        return inputs_embeds, all_position_embeds, rope_deltas
+
+    def _preprocess_decoder(
+        self,
+        input_ids: torch.LongTensor = None,
+        cache_position: torch.LongTensor = None,
+    ):
+        if self.rbln_config.batch_size != cache_position.shape[0]:
+            raise RuntimeError(
+                f"Cache position size mismatch: got {cache_position.shape[0]}, expected {self.rbln_config.batch_size}."
+            )
+
+        inputs_embeds = self.embed_tokens(input_ids)
+        position_embeds = []
+        for b_idx in range(self.rbln_config.batch_size):
+            delta = cache_position[b_idx] + self.rope_deltas[b_idx]
+            position_ids = torch.arange(1).view(1, -1)
+            position_ids = position_ids.add(delta)
+            position_ids = position_ids.unsqueeze(0).expand(3, -1, -1)
+            position_embed = self._get_position_embeddings(torch.zeros(1, dtype=torch.float32), position_ids)
+            position_embeds.append(position_embed)
+
+        position_embeds = torch.cat(position_embeds, dim=1)
+
+        return inputs_embeds, position_embeds
+
+    def forward(
+        self,
+        input_ids: Optional[torch.LongTensor] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        pixel_values: Optional[torch.Tensor] = None,
+        pixel_values_videos: Optional[torch.FloatTensor] = None,
+        image_grid_thw: Optional[torch.LongTensor] = None,
+        video_grid_thw: Optional[torch.LongTensor] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        generate_idx: Optional[torch.Tensor] = None,
+        return_dict: Optional[bool] = None,
+        **kwargs,
+    ) -> RBLNDecoderOnlyOutput:
+        # Prefill
+        if cache_position is None:
+            inputs_embeds, position_embed, rope_deltas = self._preprocess_prefill(
+                input_ids,
+                attention_mask,
+                pixel_values,
+                pixel_values_videos,
+                image_grid_thw,
+                video_grid_thw,
+            )
+
+            self.rope_deltas = rope_deltas
+            batch_size = inputs_embeds.shape[0]
+
+            logits = []
+            for b_idx in range(batch_size):
+                cache_position = torch.arange(0, generate_idx[b_idx].item(), dtype=torch.int32).unsqueeze(0)
+
+                output = self.prefill_decoder(
+                    inputs_embeds=inputs_embeds[b_idx : b_idx + 1],
+                    attention_mask=attention_mask[b_idx] if attention_mask is not None else None,
+                    cache_position=cache_position,
+                    batch_idx=b_idx,
+                    position_embed=position_embed[:, b_idx : b_idx + 1],
+                )
+                logits.append(output.logits)
+            logits = torch.cat(logits, dim=0)
+
+        # Decoder
+        else:
+            inputs_embeds, position_embed = self._preprocess_decoder(input_ids, cache_position)
+            output = self.decoder(
+                inputs_embeds=inputs_embeds,
+                cache_position=cache_position,
+                position_embed=position_embed,
+            )
+            logits = output.logits
+
+        if not return_dict:
+            return logits, generate_idx
+        else:
+            return RBLNDecoderOnlyOutput(
+                logits=logits,
+                generate_idx=generate_idx,
+            )