fount-vlm-nell-02 0.3.11__py3-none-any.whl

mlx_vlm/models/aya_vision/vision.py

@@ -0,0 +1,340 @@
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+
+from ..interpolate import resize_bilinear
+from .config import VisionConfig
+
+
+def check_array_shape(arr):
+    shape = arr.shape
+
+    # Check if the shape has 4 dimensions
+    if len(shape) != 4:
+        return False
+
+    out_channels, kH, KW, _ = shape
+
+    # 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
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dims: int,
+        num_heads: int,
+        query_input_dims: Optional[int] = None,
+        key_input_dims: Optional[int] = None,
+        value_input_dims: Optional[int] = None,
+        value_dims: Optional[int] = None,
+        value_output_dims: Optional[int] = None,
+        bias: bool = True,
+    ):
+        super().__init__()
+
+        if (dims % num_heads) != 0:
+            raise ValueError(
+                "The input feature dimensions should be divisible by the "
+                f"number of heads ({dims} % {num_heads}) != 0"
+            )
+
+        query_input_dims = query_input_dims or dims
+        key_input_dims = key_input_dims or dims
+        value_input_dims = value_input_dims or key_input_dims
+        value_dims = value_dims or dims
+        value_output_dims = value_output_dims or dims
+
+        self.num_heads = num_heads
+        head_dim = dims // num_heads
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(query_input_dims, dims, bias=bias)
+        self.k_proj = nn.Linear(key_input_dims, dims, bias=bias)
+        self.v_proj = nn.Linear(value_input_dims, value_dims, bias=bias)
+        self.out_proj = nn.Linear(value_dims, value_output_dims, bias=bias)
+
+    def __call__(self, x, mask=None):
+        queries = self.q_proj(x)
+        keys = self.k_proj(x)
+        values = self.v_proj(x)
+
+        num_heads = self.num_heads
+        B, L, D = queries.shape
+        _, S, _ = keys.shape
+        queries = queries.reshape(B, L, num_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, S, num_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, S, num_heads, -1).transpose(0, 2, 1, 3)
+
+        output = mx.fast.scaled_dot_product_attention(
+            queries, keys, values, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.out_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.activation_fn = nn.GELU(approx="precise")
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size, bias=True)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size, bias=True)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = self.fc1(x)
+        x = self.activation_fn(x)
+        x = self.fc2(x)
+        return x
+
+
+class EncoderLayer(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = Attention(
+            config.hidden_size, config.num_attention_heads, bias=True
+        )
+        self.layer_norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = MLP(config)
+        self.layer_norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    def __call__(self, x: mx.array, mask: Optional[mx.array] = None) -> mx.array:
+        r = self.self_attn(self.layer_norm1(x), mask)
+        h = x + r
+        r = self.mlp(self.layer_norm2(h))
+        return h + r
+
+
+class Encoder(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.layers = [EncoderLayer(config) for _ in range(config.num_hidden_layers)]
+
+    def __call__(
+        self,
+        x: mx.array,
+        output_hidden_states: Optional[bool] = None,
+        mask: Optional[mx.array] = None,
+    ) -> mx.array:
+        encoder_states = (x,) if output_hidden_states else None
+        h = x
+        for l in self.layers:
+            x = l(x, mask=mask)
+            if output_hidden_states:
+                encoder_states = encoder_states + (x,)
+
+            h = x
+
+        return (h, encoder_states)
+
+
+def gaussian_blur_axis(image, sigma, axis):
+    """
+    Applies a 1D Gaussian blur along the given axis.
+    This version works for arrays with any number of dimensions.
+    """
+    radius = int(3 * sigma)
+    if radius < 1:
+        return image
+    x = mx.arange(-radius, radius + 1)
+    kernel = mx.exp(-(x**2) / (2 * sigma**2))
+    kernel = kernel / mx.sum(kernel)
+
+    # MLX doesn't have a direct apply_along_axis equivalent,
+    # so we'll implement the convolution differently based on the axis
+
+    # Helper function to apply 1D convolution along specific axis
+    def conv_1d(array, kernel, axis):
+        # Reshape kernel to broadcast along the right dimensions
+        kernel_shape = [1] * image.ndim
+        kernel_shape[axis] = len(kernel)
+        kernel_reshaped = kernel.reshape(kernel_shape)
+
+        # Pad the array
+        pad_width = [(0, 0)] * image.ndim
+        pad_width[axis] = (radius, radius)
+        padded = mx.pad(array, pad_width, mode="edge")
+
+        # Perform convolution via sliding window sum
+        result = mx.zeros_like(array)
+        slices = [slice(None)] * padded.ndim
+
+        for i in range(2 * radius + 1):
+            slices[axis] = slice(i, i + array.shape[axis])
+            result = result + padded[tuple(slices)] * kernel_reshaped
+
+        return result
+
+    return conv_1d(image, kernel, axis)
+
+
+class VisionEmbeddings(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.patch_embedding = nn.Conv2d(
+            config.num_channels,
+            config.hidden_size,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+
+    @staticmethod
+    def resize_positional_embeddings(
+        positional_embeddings: mx.array,
+        spatial_shapes: mx.array,
+        max_length: int,
+    ) -> mx.array:
+        """
+        Resize positional embeddings to image-specific size and pad to a fixed size.
+
+        Args:
+            positional_embeddings (`torch.Tensor`):
+                Position embeddings of shape (height, width, embed_dim)
+            spatial_shapes (`torch.LongTensor`):
+                Spatial shapes of shape (batch_size, 2) to resize the positional embeddings to
+            max_length (`int`):
+                Maximum length of the positional embeddings to pad resized positional embeddings to
+
+        Returns:
+            `torch.Tensor`: Embeddings of shape (batch_size, max_length, embed_dim)
+        """
+        batch_size = spatial_shapes.shape[0]
+        embed_dim = positional_embeddings.shape[-1]
+        source_dtype = positional_embeddings.dtype
+
+        resulted_positional_embeddings = mx.zeros(
+            (batch_size, max_length, embed_dim)
+        ).astype(source_dtype)
+
+        # (height, width, embed_dim) -> (1, embed_dim, height, width) for interpolation
+        positional_embeddings = positional_embeddings.transpose(2, 0, 1).reshape(
+            1, embed_dim, -1
+        )
+
+        # Upcast to float32 on CPU because antialias is not supported for bfloat16/float16 on CPU
+        if positional_embeddings.device.type == "cpu":
+            positional_embeddings = positional_embeddings.astype(mx.float32)
+
+        for i in range(batch_size):
+            # (1, dim, height, width) -> (1, dim, target_height, target_width)
+            height, width = spatial_shapes[i]
+            # Then upsample width dimension
+            resized_embeddings = resize_bilinear(
+                positional_embeddings,
+                (height, width),
+                align_corners=False,
+                antialias=True,
+            )
+
+            # (1, dim, target_height, target_width) -> (target_height * target_width, dim)
+            resized_embeddings = resized_embeddings.reshape(
+                embed_dim, height * width
+            ).transpose(0, 1)
+
+            # Cast to original dtype
+            resized_embeddings = resized_embeddings.astype(source_dtype)
+
+            resulted_positional_embeddings[i, : height * width] = resized_embeddings
+            resulted_positional_embeddings[i, height * width :] = resized_embeddings[0]
+
+        return resulted_positional_embeddings
+
+    def __call__(
+        self, x: mx.array, spatial_shapes: Optional[mx.array] = None
+    ) -> mx.array:
+        batch_size = x.shape[0]
+        patch_embeddings = self.patch_embedding(x)
+        patch_embeddings = mx.flatten(patch_embeddings, start_axis=1, end_axis=2)
+        if spatial_shapes is None:
+            position_ids = mx.array(np.arange(self.num_positions)[None, :])
+            embeddings = patch_embeddings
+            embeddings += self.position_embedding(position_ids)
+
+        else:
+            # Get positional resized and padded positional embeddings
+            positional_embeddings = self.position_embedding.weight.reshape(
+                self.position_embedding_size, self.position_embedding_size, -1
+            )
+
+            resized_positional_embeddings = self.resize_positional_embeddings(
+                positional_embeddings, spatial_shapes, max_length=x.shape[1]
+            )
+
+            # Add positional embeddings to patch embeddings
+            embeddings = patch_embeds + resized_positional_embeddings
+        return embeddings
+
+
+class SigLipVisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+
+        self.embeddings = VisionEmbeddings(config)
+        self.encoder = Encoder(config)
+        self.post_layernorm = nn.LayerNorm(config.hidden_size)
+
+    def __call__(
+        self,
+        x: mx.array,
+        spatial_shapes: mx.array,
+        output_hidden_states: Optional[bool] = None,
+    ) -> mx.array:
+        x = self.embeddings(x, spatial_shapes)
+        x = x.astype(self.embeddings.patch_embedding.weight.dtype)
+        encoder_outputs = self.encoder(
+            x=x, output_hidden_states=output_hidden_states, mask=None
+        )
+        pooler_output = self.post_layernorm(encoder_outputs[0])
+        return pooler_output, x, encoder_outputs[-1]
+
+
+class VisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.model_type = config.model_type
+        if self.model_type not in ["siglip_vision_model"]:
+            raise ValueError(f"Unsupported model type: {self.model_type}")
+
+        self.vision_model = SigLipVisionModel(config)
+
+    def __call__(
+        self,
+        x: mx.array,
+        spatial_shapes: Optional[mx.array] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> mx.array:
+        return self.vision_model(x, spatial_shapes, output_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_embedding.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, 1)
+            else:
+                sanitized_weights[k] = v
+
+        return sanitized_weights

mlx_vlm/models/base.py

@@ -0,0 +1,356 @@
+import inspect
+import math
+from abc import abstractmethod
+from dataclasses import dataclass
+from typing import Dict, List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+from mlx_lm.models.base import create_attention_mask, scaled_dot_product_attention
+from PIL import Image
+
+
+@dataclass
+class LanguageModelOutput:
+    logits: mx.array
+    hidden_states: Optional[List[mx.array]] = None
+    cross_attention_states: Optional[List[mx.array]] = None
+    encoder_outputs: Optional[List[mx.array]] = None
+
+
+@dataclass
+class InputEmbeddingsFeatures:
+    inputs_embeds: mx.array
+    attention_mask_4d: Optional[mx.array] = None
+    visual_pos_masks: Optional[mx.array] = None
+    deepstack_visual_embeds: Optional[mx.array] = None
+    per_layer_inputs: Optional[mx.array] = None
+    cross_attention_states: Optional[mx.array] = None
+    cross_attention_mask: Optional[mx.array] = None
+    full_text_row_masked_out_mask: Optional[mx.array] = None
+    decoder_inputs_embeds: Optional[mx.array] = None
+    attention_mask: Optional[mx.array] = None  # For encoder-decoder models
+
+    def to_dict(self):
+        return {
+            "inputs_embeds": self.inputs_embeds,
+            "attention_mask_4d": self.attention_mask_4d,
+            "visual_pos_masks": self.visual_pos_masks,
+            "deepstack_visual_embeds": self.deepstack_visual_embeds,
+            "per_layer_inputs": self.per_layer_inputs,
+            "cross_attention_states": self.cross_attention_states,
+            "cross_attention_mask": self.cross_attention_mask,
+            "full_text_row_masked_out_mask": self.full_text_row_masked_out_mask,
+            "decoder_inputs_embeds": self.decoder_inputs_embeds,
+            "attention_mask": self.attention_mask,
+        }
+
+
+@dataclass
+class BaseModelConfig:
+    @classmethod
+    def from_dict(cls, params):
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )
+
+    def to_dict(self):
+        return {k: v for k, v in self.__dict__.items() if v is not None}
+
+
+class BaseImageProcessor:
+    """
+    Base image processor class. Subclasses should implement preprocess().
+    Transformers imports are deferred to __init__ for faster module loading.
+    """
+
+    def __init__(
+        self,
+        image_mean=(0.5, 0.5, 0.5),
+        image_std=(0.5, 0.5, 0.5),
+        size=(384, 384),
+        crop_size: Dict[str, int] = None,
+        resample=None,
+        rescale_factor=1 / 255,
+        data_format=None,
+    ):
+        from transformers.image_processing_utils import get_size_dict
+        from transformers.image_utils import ChannelDimension, PILImageResampling
+
+        if resample is None:
+            resample = PILImageResampling.BICUBIC
+        if data_format is None:
+            data_format = ChannelDimension.FIRST
+
+        crop_size = (
+            crop_size if crop_size is not None else {"height": 384, "width": 384}
+        )
+        crop_size = get_size_dict(
+            crop_size, default_to_square=True, param_name="crop_size"
+        )
+
+        self.image_mean = image_mean
+        self.image_std = image_std
+        self.size = size
+        self.resample = resample
+        self.rescale_factor = rescale_factor
+        self.data_format = data_format
+        self.crop_size = crop_size
+
+    def rescale(
+        self,
+        image,
+        scale: float,
+        input_data_format: str = "channels_first",
+    ):
+        """Rescale an image by a scale factor."""
+        return image * scale
+
+    def normalize(
+        self,
+        image,
+        mean,
+        std,
+        input_data_format: str = "channels_first",
+    ):
+        """Normalize an image with mean and std."""
+        import numpy as np
+
+        mean = np.array(mean, dtype=image.dtype)
+        std = np.array(std, dtype=image.dtype)
+
+        if input_data_format == "channels_first":
+            # Image shape: [C, H, W]
+            mean = mean[:, None, None]
+            std = std[:, None, None]
+        else:
+            # Image shape: [H, W, C]
+            pass  # mean and std are already in correct shape
+
+        return (image - mean) / std
+
+    @abstractmethod
+    def preprocess(self, images):
+        pass
+
+
+def expand2square(pil_img, background_color):
+    width, height = pil_img.size
+    if width == height:
+        return pil_img
+    elif width > height:
+        result = Image.new(pil_img.mode, (width, width), background_color)
+        result.paste(pil_img, (0, (width - height) // 2))
+        return result
+    else:
+        result = Image.new(pil_img.mode, (height, height), background_color)
+        result.paste(pil_img, ((height - width) // 2, 0))
+        return result
+
+
+def check_array_shape(arr):
+    shape = arr.shape
+
+    # Check if the shape has 4 dimensions
+    if len(shape) == 4:
+        out_channels, kH, KW, _ = shape
+        # 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
+    # Check if the shape has 3 dimensions
+    elif len(shape) == 3:
+        _, kW, out_channels = shape
+        # Check if out_channels is the largest
+        if kW >= out_channels:
+            return True
+        else:
+            return False
+    else:
+        return False
+
+
+def check_activation_stats(name, tensor):
+    """Helper function to check for anomalies and log stats."""
+
+    print(f"--- Activation Stats: {name} ---")
+    # Check for NaNs/Infs
+    has_nan = mx.isnan(tensor).any()
+    has_inf = mx.isinf(tensor).any()
+    if has_nan:
+        print(f"WARNING: Found NaN in {name}")
+    if has_inf:
+        print(f"WARNING: Found Inf in {name}")
+
+    # Calculate and print stats (ensure computation happens)
+    min_val = mx.min(tensor).item()
+    max_val = mx.max(tensor).item()
+    mean_val = mx.mean(tensor).item()
+    std_val = mx.std(tensor).item()
+    print(f"  Shape: {tensor.shape}")
+    print(f"  Min: {min_val:.4f}, Max: {max_val:.4f}")
+    print(f"  Mean: {mean_val:.4f}, Std: {std_val:.4f}")
+    print("-" * (len(name) + 24))
+
+
+def pixel_shuffle(input_tensor, shuffle_ratio):
+    # input_tensor: [batch_size, num_patches, channels]
+    batch_size, num_patches, channels = input_tensor.shape
+    patch_size = int(math.sqrt(num_patches))
+
+    input_tensor = input_tensor.reshape(batch_size, patch_size, patch_size, -1)
+    batch_size, height, width, channels = input_tensor.shape
+
+    reshaped_tensor = input_tensor.reshape(
+        batch_size, height, int(width * shuffle_ratio), int(channels / shuffle_ratio)
+    )
+    reshaped_tensor = reshaped_tensor.transpose(0, 2, 1, 3)
+
+    reshaped_tensor = reshaped_tensor.reshape(
+        batch_size,
+        int(height * shuffle_ratio),
+        int(width * shuffle_ratio),
+        int(channels / (shuffle_ratio**2)),
+    )
+    reshaped_tensor = reshaped_tensor.transpose(0, 2, 1, 3)
+
+    output_tensor = reshaped_tensor.reshape(batch_size, -1, reshaped_tensor.shape[-1])
+    return output_tensor
+
+
+def interpolate(pos_embed, size, mode="cubic", align_corners=False):
+    """
+    MLX implementation of PyTorch's F.interpolate with bicubic mode
+
+    Args:
+        pos_embed: MLX array with shape [B, C, H_src, W_src] or [C, H_src, W_src]
+        size: Tuple (H_dst, W_dst) - target size
+        align_corners: Boolean - whether to align corners
+
+    Returns:
+        Interpolated array with shape [B, C, H_dst, W_dst] or [C, H_dst, W_dst]
+    """
+    # Handle different input shapes
+    input_dim = pos_embed.ndim
+    original_shape = pos_embed.shape
+
+    if input_dim == 3:
+        # [C, H, W] -> [1, C, H, W]
+        pos_embed = pos_embed.reshape(1, *original_shape)
+
+    # Get source dimensions
+    h_src, w_src = pos_embed.shape[-2:]
+    h_dst, w_dst = size
+
+    # Calculate scale factors
+    scale_h = h_dst / h_src
+    scale_w = w_dst / w_src
+
+    # Create upsampler
+    upsampler = nn.Upsample(
+        scale_factor=(scale_h, scale_w), mode=mode, align_corners=align_corners
+    )
+
+    # Apply upsampling
+    result = upsampler(pos_embed)
+
+    # Return in the original dimension format
+    if input_dim == 3:
+        return result.reshape(original_shape[0], *size)
+    return result
+
+
+@mx.compile
+def chunked_attention(
+    queries: mx.array,
+    keys: mx.array,
+    values: mx.array,
+    scale: float,
+    chunk_size: int,
+) -> mx.array:
+
+    L = queries.shape[2]
+
+    outputs = []
+    for i in range(0, L, chunk_size):
+        end_idx = min(i + chunk_size, L)
+        q_chunk = queries[:, :, i:end_idx, :]  # (B, n_heads, chunk, head_dim)
+
+        chunk_output = mx.fast.scaled_dot_product_attention(
+            q_chunk, keys, values, scale=scale
+        )
+
+        outputs.append(chunk_output)
+
+    return mx.concatenate(outputs, axis=2)  # (B, n_heads, L, head_dim)
+
+
+def install_auto_processor_patch(target_model_types, processor_cls):
+    """
+    Install a composable patch on transformers.AutoProcessor.from_pretrained
+
+    Args:
+        target_model_types (Union[str, List[str]]): Model types to intercept.
+        processor_cls (type): Processor class exposing `from_pretrained`.
+
+    Returns:
+        The previous `AutoProcessor.from_pretrained` for reference.
+    """
+    from transformers import AutoProcessor as _HF_AutoProcessor
+
+    if isinstance(target_model_types, str):
+        target_model_types = [target_model_types]
+    target_model_types = {t.lower() for t in target_model_types}
+
+    previous_from_pretrained = _HF_AutoProcessor.from_pretrained
+
+    @classmethod
+    def _patched_auto_processor_from_pretrained(
+        cls, pretrained_model_name_or_path, **kwargs
+    ):
+        import json as _json
+        from pathlib import Path
+
+        try:
+            model_path = Path(pretrained_model_name_or_path)
+            is_local = model_path.exists() and model_path.is_dir()
+
+            cfg = {}
+            if is_local:
+                config_path = model_path / "config.json"
+                if config_path.exists():
+                    with open(config_path, "r", encoding="utf-8") as f:
+                        cfg = _json.load(f)
+            else:
+                try:
+                    from huggingface_hub import hf_hub_download
+
+                    cfg_path = hf_hub_download(
+                        pretrained_model_name_or_path, "config.json"
+                    )
+                    with open(cfg_path, "r", encoding="utf-8") as f:
+                        cfg = _json.load(f)
+                except Exception:
+                    cfg = {}
+
+            model_type = str(cfg.get("model_type", "")).lower()
+            if model_type in target_model_types:
+                return processor_cls.from_pretrained(
+                    pretrained_model_name_or_path, **kwargs
+                )
+        except Exception:
+            # On any failure, fall back to previous behavior
+            pass
+
+        # Chain to the prior from_pretrained (which may already be patched)
+        return previous_from_pretrained.__func__(
+            cls, pretrained_model_name_or_path, **kwargs
+        )
+
+    _HF_AutoProcessor.from_pretrained = _patched_auto_processor_from_pretrained
+    return previous_from_pretrained