nexaai 1.0.4rc10__py3-none-macosx_11_0_arm64.whl

nexaai/mlx_backend/vlm/modeling/models/qwen2_vl/vision.py

@@ -0,0 +1,312 @@
+import inspect
+from dataclasses import dataclass
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from .config import VisionConfig
+
+
+def check_array_shape(arr):
+    shape = arr.shape
+
+    # Check if the shape has 4 dimensions
+    if len(shape) not in [4, 5]:
+        return False
+
+    B, out_channels, kH, KW, t = shape
+
+    if t == 3:
+        return True
+
+    # Check if out_channels is the largest, and kH and KW are the same
+    if (out_channels >= kH) and (out_channels >= KW) and (kH == KW):
+        return True
+    else:
+        return False
+
+
+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 mx.concatenate([-x2, x1], axis=-1)
+
+
+def apply_rotary_pos_emb_vision(tensor, freqs) -> mx.array:
+    orig_dtype = tensor.dtype
+
+    cos = mx.cos(freqs)
+    sin = mx.sin(freqs)
+
+    cos = mx.expand_dims(cos, axis=1)  # Equivalent to unsqueeze(1)
+    cos = mx.tile(cos, (1, 1, 2))  # Equivalent to repeat(1, 1, 2)
+    cos = mx.expand_dims(cos, axis=0)  # Equivalent to [None, ...]
+
+    sin = mx.expand_dims(sin, axis=1)  # Equivalent to unsqueeze(1)
+    sin = mx.tile(sin, (1, 1, 2))  # Equivalent to repeat(1, 1, 2)
+    sin = mx.expand_dims(sin, axis=0)  # Equivalent to [None, ...]
+
+    output = (tensor * cos) + (rotate_half(tensor) * sin)
+    return output.astype(orig_dtype)
+
+
+class VisionRotaryEmbedding(nn.Module):
+    def __init__(self, dim: int, theta: float = 10000.0) -> None:
+        super().__init__()
+        self.dim = dim
+        self.theta = theta
+
+    def __call__(self, seqlen: int) -> mx.array:
+        inv_freq = 1.0 / (
+            self.theta ** (mx.arange(0, self.dim, 2, dtype=mx.float32) / self.dim)
+        )
+        seq = mx.arange(seqlen.tolist(), dtype=inv_freq.dtype)
+        freqs = mx.outer(seq, 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 __call__(self, hidden_states: mx.array) -> mx.array:
+        hidden_states = hidden_states.reshape(
+            -1,
+            self.in_channels,
+            self.temporal_patch_size,
+            self.patch_size,
+            self.patch_size,
+        ).moveaxis(1, 4)
+
+        hidden_states = self.proj(hidden_states)
+        hidden_states = hidden_states.reshape(-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 = nn.LayerNorm(context_dim, eps=1e-6)
+        self.mlp = [
+            nn.Linear(self.hidden_size, self.hidden_size),
+            nn.GELU(),
+            nn.Linear(self.hidden_size, dim),
+        ]
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = self.ln_q(x).reshape(-1, self.hidden_size)
+        for layer in self.mlp:
+            x = layer(x)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(self, dim: int, num_heads: int = 16) -> None:
+        super().__init__()
+        self.num_heads = num_heads
+        self.head_dim = head_dim = dim // num_heads
+        self.scale = head_dim**-0.5
+        self.qkv = nn.Linear(dim, dim * 3, bias=True)
+        self.proj = nn.Linear(dim, dim)
+
+    def __call__(
+        self, x: mx.array, cu_seqlens: mx.array, rotary_pos_emb: mx.array = None
+    ) -> mx.array:
+        seq_length = x.shape[0]
+        qkv = (
+            self.qkv(x).reshape(seq_length, 3, self.num_heads, -1).transpose(1, 0, 2, 3)
+        )
+        q, k, v = mx.split(qkv, 3)
+
+        q = apply_rotary_pos_emb_vision(mx.expand_dims(q, 0), rotary_pos_emb)[0]
+        k = apply_rotary_pos_emb_vision(mx.expand_dims(k, 0), rotary_pos_emb)[0]
+        attention_mask = mx.ones((1, seq_length, seq_length), dtype=x.dtype)
+
+        for i in range(1, len(cu_seqlens)):
+            start = int(cu_seqlens[i - 1])
+            end = int(cu_seqlens[i])
+            attention_mask[start:end, start:end] = 0
+
+        q = q.transpose(0, 2, 1, 3)
+        k = k.transpose(0, 2, 1, 3)
+        v = v.transpose(0, 2, 1, 3)
+
+        output = mx.fast.scaled_dot_product_attention(
+            q, k, v, scale=self.scale, mask=attention_mask
+        )
+        output = output.transpose(0, 2, 1, 3)
+        output = output.reshape(seq_length, -1)
+        return self.proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.activation_fn = nn.GELU(approx="fast")
+        self.fc1 = nn.Linear(dim, hidden_dim)
+        self.fc2 = nn.Linear(hidden_dim, dim)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = self.activation_fn(self.fc1(x))
+        x = self.fc2(x)
+        return x
+
+
+class Qwen2VLVisionBlock(nn.Module):
+    def __init__(self, config: VisionConfig) -> None:
+        super().__init__()
+        self.norm1 = nn.LayerNorm(config.embed_dim, eps=1e-6)
+        self.norm2 = nn.LayerNorm(config.embed_dim, eps=1e-6)
+        mlp_hidden_dim = int(config.embed_dim * config.mlp_ratio)
+
+        self.attn = Attention(dim=config.embed_dim, num_heads=config.num_heads)
+        self.mlp = MLP(dim=config.embed_dim, hidden_dim=mlp_hidden_dim)
+
+    def __call__(self, hidden_states, cu_seqlens, rotary_pos_emb) -> mx.array:
+        hidden_states = hidden_states + self.attn(
+            self.norm1(hidden_states),
+            cu_seqlens=cu_seqlens,
+            rotary_pos_emb=rotary_pos_emb,
+        )
+        hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
+        return hidden_states
+
+
+class VisionModel(nn.Module):
+
+    def __init__(self, config: VisionConfig) -> None:
+        super().__init__()
+        self.config = config
+        self.model_type = config.model_type
+        if self.model_type != "qwen2_vl":
+            raise ValueError(f"Unsupported model type: {self.model_type}")
+        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 = [Qwen2VLVisionBlock(config) for _ in range(config.depth)]
+        self.merger = PatchMerger(dim=config.hidden_size, context_dim=config.embed_dim)
+
+    def rot_pos_emb(self, grid_thw):
+        pos_ids = []
+
+        for t, h, w in grid_thw:
+            h, w = int(h), int(w)  # Ensure h and w are integers
+            hpos_ids = mx.expand_dims(mx.arange(h), 1)
+            hpos_ids = mx.repeat(hpos_ids, w, axis=1)
+            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 = mx.transpose(hpos_ids, (0, 2, 1, 3))
+            hpos_ids = hpos_ids.flatten()
+
+            wpos_ids = mx.expand_dims(mx.arange(w), 0)
+            wpos_ids = mx.repeat(wpos_ids, h, axis=0)
+            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 = mx.transpose(wpos_ids, (0, 2, 1, 3))
+            wpos_ids = wpos_ids.flatten()
+
+            stacked_pos_ids = mx.stack([hpos_ids, wpos_ids], axis=-1)
+            pos_ids.append(mx.tile(stacked_pos_ids, (t, 1)))
+
+        pos_ids = mx.concatenate(pos_ids, axis=0)
+        max_grid_size = mx.max(grid_thw[:, 1:])
+        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+
+        rotary_pos_emb_full = rotary_pos_emb_full[pos_ids]
+
+        return rotary_pos_emb_full.reshape(pos_ids.shape[0], -1)
+
+    def __call__(
+        self,
+        hidden_states: mx.array,
+        grid_thw: mx.array,
+        output_hidden_states: Optional[bool] = None,
+    ) -> mx.array:
+
+        hidden_states = self.patch_embed(hidden_states)
+        rotary_pos_emb = self.rot_pos_emb(grid_thw)
+
+        # Assuming grid_thw has shape (batch_size, 3)
+        batch_size = grid_thw.shape[0]
+
+        # Calculate cu_seqlens for each item in the batch
+        cu_seqlens = []
+        for i in range(batch_size):
+            seq_len = grid_thw[i, 1] * grid_thw[i, 2]
+            cu_seqlens.append(mx.repeat(seq_len, grid_thw[i, 0]))
+
+        # Concatenate the cu_seqlens for all items in the batch
+        cu_seqlens = mx.concatenate(cu_seqlens)
+
+        cu_seqlens = mx.cumsum(cu_seqlens.astype(mx.int32), axis=0)
+        cu_seqlens = mx.pad(cu_seqlens, (1, 0), mode="constant", constant_values=0)
+
+        encoder_states = (hidden_states,) if output_hidden_states else None
+
+        for blk in self.blocks:
+            hidden_states = blk(
+                hidden_states, cu_seqlens=cu_seqlens, rotary_pos_emb=rotary_pos_emb
+            )
+            if output_hidden_states:
+                encoder_states = encoder_states + (hidden_states,)
+
+        return self.merger(hidden_states)
+
+    def sanitize(self, weights):
+        sanitized_weights = {}
+        for k, v in weights.items():
+            if "position_ids" in k:
+                # Remove unused position_ids
+                continue
+            elif "patch_embed.proj.weight" in k:
+                # PyTorch conv2d weight tensors have shape:
+                #   [out_channels, in_channels, kH, KW]
+                # MLX conv2d expects the weight be of shape:
+                #   [out_channels, kH, KW, in_channels]
+                if check_array_shape(v):
+                    sanitized_weights[k] = v
+                else:
+                    sanitized_weights[k] = v.transpose(0, 2, 3, 4, 1)
+            else:
+                sanitized_weights[k] = v
+
+        return sanitized_weights

nexaai/mlx_backend/vlm/modeling/models/smolvlm/__init__.py

@@ -0,0 +1,8 @@
+from .smolvlm import (
+    LanguageModel,
+    Model,
+    ModelConfig,
+    TextConfig,
+    VisionConfig,
+    VisionModel,
+)

nexaai/mlx_backend/vlm/modeling/models/smolvlm/smolvlm.py

@@ -0,0 +1,62 @@
+import mlx.core as mx
+import numpy as np
+
+from ..idefics3 import LanguageModel
+from ..idefics3 import Model as Idefics3Model
+from ..idefics3 import ModelConfig, TextConfig, VisionConfig, VisionModel
+
+
+class Model(Idefics3Model):
+    def _prepare_inputs_for_multimodal(self, image_features, inputs_embeds, input_ids):
+        # Assumes bs == 1
+
+        B, T, D_text = inputs_embeds.shape
+        N, S, D_img = image_features.shape
+
+        image_offset = 0
+        cur_embeds = inputs_embeds[0]
+
+        # Find positions of <image> tokens in the text
+        image_token_index = self.config.image_token_index
+        image_positions = np.where(input_ids == image_token_index)[1].tolist()
+        num_image_tokens = len(image_positions)
+
+        # If no <image> => text-only
+        if num_image_tokens == 0:
+            empty_slice = image_features[0][:0, :]  # shape (0, D)
+            return mx.concatenate([cur_embeds, empty_slice], axis=0)
+
+        # Typically, if each image is S embeddings, we expect the total # of <image> tokens
+        # in this sample to be multiple of S => each group of S tokens = 1 image
+        if num_image_tokens % S != 0:
+            raise ValueError(
+                f"Input has {num_image_tokens} <image> tokens, not a multiple of S={S}. "
+                "Cannot map them to blocks of shape (S, D)."
+            )
+
+        chunks = [image_positions[i : i + S] for i in range(0, num_image_tokens, S)]
+
+        segments = []
+        text_start = 0
+
+        # For each chunk (each chunk => 1 image)
+        for chunk in chunks:
+            cur_block = image_features[image_offset]
+            image_offset += 1
+
+            # We'll iterate over the S positions in ascending order
+            for i_s, pos in enumerate(chunk):
+                if pos > text_start:
+                    segments.append(cur_embeds[text_start:pos])
+                # Then add one row from cur_block => shape (1, D)
+                row_of_block = cur_block[i_s : i_s + 1, :]
+                segments.append(row_of_block)
+                text_start = pos + 1
+
+        # leftover text after the final <image> token
+        if text_start < T:
+            segments.append(cur_embeds[text_start:])
+
+        # cat them into a single (T_b, D) tensor
+        merged_sample = mx.concatenate(segments, axis=0)
+        return mx.expand_dims(merged_sample, axis=0)

nexaai/mlx_backend/vlm/modeling/processing_qwen2_5_vl.py

@@ -0,0 +1,209 @@
+# Copied from transformers. Removed video-related code.
+
+from typing import Optional, Union
+
+import numpy as np
+
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.image_utils import ImageInput
+from transformers.processing_utils import ImagesKwargs, MultiModalData, ProcessingKwargs, ProcessorMixin, Unpack, VideosKwargs
+from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
+
+
+class Qwen2_5_VLImagesKwargs(ImagesKwargs):
+    min_pixels: Optional[int]
+    max_pixels: Optional[int]
+    patch_size: Optional[int]
+    temporal_patch_size: Optional[int]
+    merge_size: Optional[int]
+
+
+class Qwen2_5_VLProcessorKwargs(ProcessingKwargs, total=False):
+    images_kwargs: Qwen2_5_VLImagesKwargs
+    _defaults = {
+        "text_kwargs": {
+            "padding": False,
+            "return_mm_token_type_ids": False,
+        },
+    }
+
+
+class Qwen2_5_VLProcessor(ProcessorMixin):
+    r"""
+    Constructs a Qwen2.5-VL processor which wraps a Qwen2.5-VL image processor and a Qwen2 tokenizer into a single processor.
+    [`Qwen2_5_VLProcessor`] offers all the functionalities of [`Qwen2VLImageProcessor`] and [`Qwen2TokenizerFast`]. See the
+    [`~Qwen2_5_VLProcessor.__call__`] and [`~Qwen2_5_VLProcessor.decode`] for more information.
+    Args:
+        image_processor ([`Qwen2VLImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`Qwen2TokenizerFast`], *optional*):
+            The tokenizer is a required input.
+        chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
+            in a chat into a tokenizable string.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+
+    image_processor_class = "AutoImageProcessor"
+    tokenizer_class = ("Qwen2Tokenizer", "Qwen2TokenizerFast")
+
+    def __init__(self, image_processor=None, tokenizer=None, chat_template=None, **kwargs):
+        self.image_token = "<|image_pad|>" if not hasattr(tokenizer, "image_token") else tokenizer.image_token
+        self.image_token_id = (
+            tokenizer.image_token_id
+            if getattr(tokenizer, "image_token_id", None)
+            else tokenizer.convert_tokens_to_ids(self.image_token)
+        )
+        super().__init__(image_processor, tokenizer, chat_template=chat_template)
+
+    def __call__(
+        self,
+        images: ImageInput = None,
+        text: Union[TextInput, PreTokenizedInput, list[TextInput], list[PreTokenizedInput]] = None,
+        **kwargs: Unpack[Qwen2_5_VLProcessorKwargs],
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwrags` arguments to
+        Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`] if `vision_infos` is not `None`.
+
+        Args:
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `list[PIL.Image.Image]`, `list[np.ndarray]`, `list[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            text (`str`, `list[str]`, `list[list[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+            - **image_grid_thw** -- List of image 3D grid in LLM. Returned when `images` is not `None`.
+        """
+        output_kwargs = self._merge_kwargs(
+            Qwen2_5_VLProcessorKwargs,
+            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
+            **kwargs,
+        )
+
+        image_inputs = {}
+        if images is not None:
+            image_inputs = self.image_processor(images=images, **output_kwargs["images_kwargs"])
+            image_grid_thw = image_inputs["image_grid_thw"]
+
+        if not isinstance(text, list):
+            text = [text]
+
+        text = text.copy()  # below lines change text in-place
+        if images is not None:
+            merge_length = self.image_processor.merge_size**2
+            index = 0
+            for i in range(len(text)):
+                while self.image_token in text[i]:
+                    num_image_tokens = image_grid_thw[index].prod() // merge_length
+                    text[i] = text[i].replace(self.image_token, "<|placeholder|>" * num_image_tokens, 1)
+                    index += 1
+                text[i] = text[i].replace("<|placeholder|>", self.image_token)
+
+        return_tensors = output_kwargs["text_kwargs"].pop("return_tensors", None)
+        return_mm_token_type_ids = output_kwargs["text_kwargs"].pop("return_mm_token_type_ids", None)
+        text_inputs = self.tokenizer(text, **output_kwargs["text_kwargs"])
+        self._check_special_mm_tokens(text, text_inputs, modalities=["image"])
+
+        if return_mm_token_type_ids:
+            array_ids = np.array(text_inputs["input_ids"])
+            mm_token_type_ids = np.zeros_like(text_inputs["input_ids"])
+            mm_token_type_ids[array_ids == self.image_token_id] = 1
+            text_inputs["mm_token_type_ids"] = mm_token_type_ids.tolist()
+
+        return BatchFeature(data={**text_inputs, **image_inputs}, tensor_type=return_tensors)
+
+    def _get_num_multimodal_tokens(self, image_sizes=None, **kwargs):
+        """
+        Computes the number of placeholder tokens needed for multimodal inputs with the given sizes.
+        Args:
+            image_sizes (`list[list[int]]`, *optional*):
+                The input sizes formatted as (height, width) per each image.
+        Returns:
+            `MultiModalData`: A `MultiModalData` object holding number of tokens per each of the provided
+            input modalities, along with other useful data.
+        """
+
+        vision_data = {}
+        if image_sizes is not None:
+            images_kwargs = Qwen2_5_VLProcessorKwargs._defaults.get("images_kwargs", {})
+            images_kwargs.update(kwargs)
+            merge_size = images_kwargs.get("merge_size", None) or self.image_processor.merge_size
+
+            num_image_patches = [
+                self.image_processor.get_number_of_image_patches(*image_size, images_kwargs)
+                for image_size in image_sizes
+            ]
+            num_image_tokens = [(num_patches // merge_size**2) for num_patches in num_image_patches]
+            vision_data.update({"num_image_tokens": num_image_tokens, "num_image_patches": num_image_patches})
+
+        return MultiModalData(**vision_data)
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to Qwen2TokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to Qwen2TokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    def post_process_image_text_to_text(
+        self, generated_outputs, skip_special_tokens=True, clean_up_tokenization_spaces=False, **kwargs
+    ):
+        """
+        Post-process the output of the model to decode the text.
+
+        Args:
+            generated_outputs (`torch.Tensor` or `np.ndarray`):
+                The output of the model `generate` function. The output is expected to be a tensor of shape `(batch_size, sequence_length)`
+                or `(sequence_length,)`.
+            skip_special_tokens (`bool`, *optional*, defaults to `True`):
+                Whether or not to remove special tokens in the output. Argument passed to the tokenizer's `batch_decode` method.
+            clean_up_tokenization_spaces (`bool`, *optional*, defaults to `False`):
+                Whether or not to clean up the tokenization spaces. Argument passed to the tokenizer's `batch_decode` method.
+            **kwargs:
+                Additional arguments to be passed to the tokenizer's `batch_decode method`.
+
+        Returns:
+            `list[str]`: The decoded text.
+        """
+        return self.tokenizer.batch_decode(
+            generated_outputs,
+            skip_special_tokens=skip_special_tokens,
+            clean_up_tokenization_spaces=clean_up_tokenization_spaces,
+            **kwargs,
+        )
+
+    @property
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        names_from_processor = list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
+        return names_from_processor
+
+
+__all__ = ["Qwen2_5_VLProcessor"]