nexaai 1.0.4rc10__py3-none-macosx_11_0_arm64.whl

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

@@ -0,0 +1,193 @@
+import glob
+import inspect
+import json
+from dataclasses import dataclass
+from pathlib import Path
+from typing import List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+from mlx_lm.utils import snapshot_download
+
+from .language import LanguageModel, TextConfig
+from .vision import VisionConfig, VisionModel
+
+
+@dataclass
+class ModelConfig:
+    text_config: TextConfig
+    vision_config: VisionConfig
+    model_type: str
+    image_token_index: int = 255036
+    max_splits_per_img: int = 12
+    downsample_factor: int = 2
+    alignment_intermediate_size: int = 28672
+    adapter_layer_norm_eps: float = 1e-06
+    vision_feature_layer: int = -1
+    vision_feature_select_strategy: str = "full"
+    eos_token_id: Optional[List[int]] = None
+
+    @classmethod
+    def from_dict(cls, params):
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )
+
+
+class AyaVisionMultiModalProjector(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.config = config
+        self.downsample_factor = config.downsample_factor
+        self.alignment_intermediate_size = getattr(
+            config, "alignment_intermediate_size", config.text_config.hidden_size
+        )
+        self.layernorm = nn.LayerNorm(
+            config.vision_config.hidden_size * (config.downsample_factor**2),
+            eps=config.adapter_layer_norm_eps,
+        )
+
+        self.linear_1 = nn.Linear(
+            config.vision_config.hidden_size * (config.downsample_factor**2),
+            self.alignment_intermediate_size,
+            bias=True,
+        )
+
+        self.act = nn.SiLU()  # SwiGLU uses SiLU activation
+
+        # For SwiGLU, project down to half size since we split intermediate dim
+        self.linear_2 = nn.Linear(
+            self.alignment_intermediate_size // 2,
+            config.text_config.hidden_size,
+            bias=True,
+        )
+
+    def __call__(self, image_features):
+        image_features = self.pixel_shuffle(image_features)
+        image_features = self.layernorm(image_features)
+        hidden_states = self.linear_1(image_features)
+
+        # Split along last dimension and apply SwiGLU
+        x, gate = mx.split(hidden_states, 2, axis=-1)
+        hidden_states = self.act(gate) * x
+
+        hidden_states = self.linear_2(hidden_states)
+        return hidden_states
+
+    def pixel_shuffle(self, image_features):  # B, S, D
+        batch_size, seq_length, feature_dim = image_features.shape
+        height = width = int(seq_length**0.5)
+        image_features = image_features.reshape(
+            image_features.shape[0], width, height, -1
+        )
+        channels = image_features.shape[-1]
+        image_features = image_features.reshape(
+            batch_size,
+            width,
+            int(height / self.downsample_factor),
+            int(channels * self.downsample_factor),
+        )
+        image_features = image_features.transpose(0, 2, 1, 3)
+        image_features = image_features.reshape(
+            batch_size,
+            int(height / self.downsample_factor),
+            int(width / self.downsample_factor),
+            -1,
+        )
+        image_features = image_features.transpose(0, 2, 1, 3)
+        return image_features
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.config = config
+        self.vision_tower = VisionModel(config.vision_config)
+        self.language_model = LanguageModel(config.text_config)
+        self.multi_modal_projector = AyaVisionMultiModalProjector(config)
+        self.vision_feature_layer = config.vision_feature_layer
+        self.vision_feature_select_strategy = config.vision_feature_select_strategy
+
+    def get_input_embeddings(
+        self,
+        input_ids: Optional[mx.array] = None,
+        pixel_values: Optional[mx.array] = None,
+        **kwargs,
+    ):
+        if pixel_values is None:
+            return self.language_model.model.embed_tokens(input_ids)
+
+        # Get the input embeddings from the language model
+        inputs_embeds = self.language_model.model.embed_tokens(input_ids)
+
+        spatial_shapes = kwargs.get("spatial_shapes", None)
+        # Get the ouptut hidden states from the vision model
+        *_, hidden_states = self.vision_tower(
+            pixel_values.transpose(0, 2, 3, 1),
+            spatial_shapes=spatial_shapes,
+            output_hidden_states=True,
+        )
+
+        # Select the hidden states from the desired layer
+        selected_image_feature = hidden_states[self.vision_feature_layer]
+
+        if self.vision_feature_select_strategy == "default":
+            selected_image_feature = selected_image_feature[:, 1:]
+        elif self.vision_feature_select_strategy == "full":
+            selected_image_feature = selected_image_feature
+        else:
+            raise ValueError(
+                "Unexpected feature selection strategy: "
+                f"{self.vision_feature_select_strategy}"
+            )
+
+        # Pass image features through the multi-modal projector
+        image_features = self.multi_modal_projector(selected_image_feature)
+
+        # Insert special image tokens in the input_ids
+        final_inputs_embeds = self._merge_input_ids_with_image_features(
+            image_features, inputs_embeds, input_ids
+        )
+        return final_inputs_embeds
+
+    def _merge_input_ids_with_image_features(
+        self, image_features, inputs_embeds, input_ids
+    ):
+        image_token_index = self.config.image_token_index
+
+        # Positions of <image> tokens in input_ids, assuming batch size is 1
+        image_positions = np.where(input_ids[0] == image_token_index)[0].tolist()
+        num_images, _, _, vision_hidden_size = image_features.shape
+
+        reshaped_image_hidden_states = image_features.reshape(-1, vision_hidden_size)
+
+        # cast to the dtype of the input_embeds to support quantized models
+        reshaped_image_hidden_states = reshaped_image_hidden_states.astype(
+            inputs_embeds.dtype
+        )
+        inputs_embeds[:, image_positions, :] = reshaped_image_hidden_states
+        return inputs_embeds
+
+    @property
+    def layers(self):
+        return self.language_model.model.layers
+
+    def __call__(
+        self,
+        input_ids: mx.array,
+        pixel_values: mx.array,
+        mask: mx.array,
+        cache=None,
+        **kwargs,
+    ):
+
+        input_embddings = self.get_input_embeddings(input_ids, pixel_values, **kwargs)
+        logits = self.language_model(
+            input_ids, cache=cache, inputs_embeds=input_embddings
+        )
+        return logits

nexaai/mlx_backend/vlm/modeling/models/aya_vision/interpolate.py

@@ -0,0 +1,186 @@
+import mlx.core as mx
+import numpy as np
+
+
+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)
+
+
+def bilinear_interpolate(image, new_height, new_width, align_corners=False):
+    """
+    Performs bilinear interpolation on an array whose spatial dimensions are the first two.
+    It supports extra dimensions (e.g. channels or batch dimensions that have been moved to the trailing axes).
+    """
+    # image is assumed to have shape (H, W, ...) where H and W are spatial dimensions.
+    H_in, W_in = image.shape[0], image.shape[1]
+
+    # Compute sampling positions in the input image.
+    if new_height == 1:
+        row_positions = mx.array([0.0])
+    else:
+        if align_corners:
+            row_positions = mx.linspace(0, H_in - 1, new_height)
+        else:
+            row_positions = (mx.arange(new_height) + 0.5) * H_in / new_height - 0.5
+
+    if new_width == 1:
+        col_positions = mx.array([0.0])
+    else:
+        if align_corners:
+            col_positions = mx.linspace(0, W_in - 1, new_width)
+        else:
+            col_positions = (mx.arange(new_width) + 0.5) * W_in / new_width - 0.5
+
+    # Compute floor and ceil indices.
+    row_floor = mx.floor(row_positions).astype(mx.int32)
+    col_floor = mx.floor(col_positions).astype(mx.int32)
+    row_ceil = row_floor + 1
+    col_ceil = col_floor + 1
+
+    row_floor = mx.clip(row_floor, 0, H_in - 1)
+    row_ceil = mx.clip(row_ceil, 0, H_in - 1)
+    col_floor = mx.clip(col_floor, 0, W_in - 1)
+    col_ceil = mx.clip(col_ceil, 0, W_in - 1)
+
+    row_weight = row_positions - row_floor  # shape (new_height,)
+    col_weight = col_positions - col_floor  # shape (new_width,)
+
+    # Use advanced indexing for gather operations
+    # Create meshgrid for coordinates
+    row_floor_grid, col_floor_grid = mx.meshgrid(row_floor, col_floor, indexing="ij")
+    row_ceil_grid, col_floor_grid = mx.meshgrid(row_ceil, col_floor, indexing="ij")
+    row_floor_grid, col_ceil_grid = mx.meshgrid(row_floor, col_ceil, indexing="ij")
+    row_ceil_grid, col_ceil_grid = mx.meshgrid(row_ceil, col_ceil, indexing="ij")
+
+    # Gather the four surrounding pixels using take_along_axis
+    # For higher dimensional arrays, we'll need to reshape and broadcast
+    extra_dims = image.ndim - 2
+
+    def gather_pixels(row_indices, col_indices):
+        # Flatten the spatial dimensions for gathering
+        flat_indices = row_indices * W_in + col_indices
+        flat_image = mx.reshape(image, (-1,) + image.shape[2:])
+        # Gather and reshape back
+        gathered = mx.take(flat_image, flat_indices.reshape(-1), axis=0)
+        return mx.reshape(gathered, (new_height, new_width) + image.shape[2:])
+
+    top_left = gather_pixels(row_floor_grid, col_floor_grid)
+    top_right = gather_pixels(row_floor_grid, col_ceil_grid)
+    bottom_left = gather_pixels(row_ceil_grid, col_floor_grid)
+    bottom_right = gather_pixels(row_ceil_grid, col_ceil_grid)
+
+    # Expand the weights to have shape (new_height, new_width, *[1]*extra_dims)
+    r_weight = row_weight.reshape(new_height, 1, *([1] * extra_dims))
+    c_weight = col_weight.reshape(1, new_width, *([1] * extra_dims))
+
+    # Perform bilinear interpolation.
+    result = (
+        (1 - r_weight) * (1 - c_weight) * top_left
+        + (1 - r_weight) * c_weight * top_right
+        + r_weight * (1 - c_weight) * bottom_left
+        + r_weight * c_weight * bottom_right
+    )
+    return result
+
+
+def resize_bilinear(image, new_size, align_corners=False, antialias=True):
+    """
+    Resizes an image (or embedding tensor) to new_size=(new_height, new_width)
+    using bilinear interpolation with MLX.
+
+    Supports:
+      - 2D: (H, W)
+      - 3D: (H, W, C)
+      - 4D: (B, C, H, W)  (assumed for typical image batches)
+    """
+    new_height, new_width = new_size
+
+    # Convert numpy arrays to MLX arrays if needed
+    if isinstance(image, np.ndarray):
+        image = mx.array(image)
+
+    if image.ndim == 2 or image.ndim == 3:
+        # Assume spatial dims are the first two.
+        resized = image
+        H_in, W_in = image.shape[:2]
+        if antialias:
+            if new_height < H_in:
+                scale_y = new_height / H_in
+                sigma_y = (1 / scale_y - 1) / 2.0  # heuristic
+                if sigma_y > 0:
+                    resized = gaussian_blur_axis(resized, sigma_y, axis=0)
+            if new_width < W_in:
+                scale_x = new_width / W_in
+                sigma_x = (1 / scale_x - 1) / 2.0
+                if sigma_x > 0:
+                    resized = gaussian_blur_axis(resized, sigma_x, axis=1)
+        resized = bilinear_interpolate(
+            resized, new_height, new_width, align_corners=align_corners
+        )
+        return resized
+
+    elif image.ndim == 4:
+        # Assume shape is (B, C, H, W) (typical PyTorch/MLX format).
+        B, C, H_in, W_in = image.shape
+        # Permute to bring spatial dims to the front: (H, W, B, C)
+        image_perm = mx.transpose(image, (2, 3, 0, 1))
+        resized = image_perm
+        if antialias:
+            if new_height < H_in:
+                scale_y = new_height / H_in
+                sigma_y = (1 / scale_y - 1) / 2.0
+                if sigma_y > 0:
+                    resized = gaussian_blur_axis(resized, sigma_y, axis=0)
+            if new_width < W_in:
+                scale_x = new_width / W_in
+                sigma_x = (1 / scale_x - 1) / 2.0
+                if sigma_x > 0:
+                    resized = gaussian_blur_axis(resized, sigma_x, axis=1)
+        resized = bilinear_interpolate(
+            resized, new_height, new_width, align_corners=align_corners
+        )
+        # Permute back to (B, C, new_height, new_width)
+        resized = mx.transpose(resized, (2, 3, 0, 1))
+        return resized
+
+    else:
+        raise ValueError("Unsupported image dimensions.")
+
+
+#

nexaai/mlx_backend/vlm/modeling/models/aya_vision/language.py

@@ -0,0 +1,233 @@
+import inspect
+from dataclasses import dataclass
+from typing import Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from ..base import (
+    LanguageModelOutput,
+    create_attention_mask,
+    scaled_dot_product_attention,
+)
+from ..cache import KVCache, RotatingKVCache
+
+
+@dataclass
+class TextConfig:
+    model_type: str
+    hidden_size: int = 8192
+    head_dim: int = 128
+    num_hidden_layers: int = 40
+    intermediate_size: int = 14336
+    num_attention_heads: int = 64
+    num_key_value_heads: int = 8
+    rope_theta: float = 50000.0
+    vocab_size: int = 256000
+    layer_norm_eps: float = 1e-05
+    logit_scale: float = 0.0625
+    attention_bias: bool = False
+    layer_norm_bias: bool = False
+    sliding_window: int = 4096
+    sliding_window_pattern: int = 4
+    max_position_embeddings: int = 4096
+
+    @classmethod
+    def from_dict(cls, params):
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )
+
+
+class Attention(nn.Module):
+    def __init__(self, config: TextConfig, layer_idx: int):
+        super().__init__()
+        self.config = config
+        self.layer_idx = layer_idx
+
+        dim = config.hidden_size
+        self.n_heads = n_heads = config.num_attention_heads
+        self.n_kv_heads = n_kv_heads = config.num_key_value_heads
+        self.head_dim = head_dim = config.head_dim
+        if (head_dim * n_heads) != dim:
+            raise ValueError(
+                f"hidden_size must be divisible by num_heads (got `hidden_size`: {dim}"
+                f" and `num_heads`: {n_heads})."
+            )
+        self.scale = head_dim**-0.5
+
+        attetion_bias = config.attention_bias
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=attetion_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attetion_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attetion_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=attetion_bias)
+
+        self.rope = nn.RoPE(head_dim, traditional=True, base=config.rope_theta)
+
+        self.use_sliding_window = (layer_idx + 1) % config.sliding_window_pattern != 0
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        queries = queries.reshape(B, L, self.n_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, self.n_kv_heads, -1).transpose(0, 2, 1, 3)
+
+        # Apply RoPE only if sliding window is enabled
+        if self.use_sliding_window:
+            if cache is None:
+                queries = self.rope(queries)
+                keys = self.rope(keys)
+            else:
+                queries = self.rope(queries, offset=cache.offset)
+                keys = self.rope(keys, offset=cache.offset)
+
+        if cache is not None:
+            keys, values = cache.update_and_fetch(keys, values)
+
+        if self.use_sliding_window and mask is not None and isinstance(mask, mx.array):
+            key_len = keys.shape[-2]
+            if mask.shape[-1] != key_len:
+                mask = mask[..., -key_len:]
+
+        output = scaled_dot_product_attention(
+            queries, keys, values, cache, scale=self.scale, mask=mask
+        )
+
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+        return self.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, dim, hidden_dim):
+        super().__init__()
+        self.gate_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+
+    def __call__(self, x):
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, config: TextConfig, layer_idx: int):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.n_heads = config.num_attention_heads
+
+        self.self_attn = Attention(config, layer_idx)
+        self.mlp = MLP(config.hidden_size, config.intermediate_size)
+        self.input_layernorm = nn.LayerNorm(
+            config.hidden_size, eps=config.layer_norm_eps, bias=config.layer_norm_bias
+        )
+        self.config = config
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[Tuple[mx.array, mx.array]] = None,
+    ) -> mx.array:
+        h = self.input_layernorm(x)
+        attn_h = self.self_attn(h, mask, cache)
+        ff_h = self.mlp(h)
+        return attn_h + ff_h + x
+
+
+class CohereModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.vocab_size = config.vocab_size
+        self.num_hidden_layers = config.num_hidden_layers
+        assert self.vocab_size > 0
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.layers = [
+            TransformerBlock(config, layer_idx=i)
+            for i in range(config.num_hidden_layers)
+        ]
+        self.norm = nn.LayerNorm(
+            config.hidden_size, eps=config.layer_norm_eps, bias=config.layer_norm_bias
+        )
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        inputs_embeds: mx.array = None,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        if inputs_embeds is None:
+            h = self.embed_tokens(inputs)
+        else:
+            h = inputs_embeds
+
+        if cache is None:
+            cache = [None] * len(self.layers)
+
+        if mask is None:
+            j = self.config.sliding_window_pattern
+            mask = create_attention_mask(h, cache[j - 1 : j])
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        return self.norm(h)
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.model_type = config.model_type
+        self.model = CohereModel(config)
+        self.config = config
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        inputs_embeds: mx.array = None,
+        mask: mx.array = None,
+        cache=None,
+    ):
+        out = self.model(inputs, inputs_embeds, mask, cache)
+        out = self.model.embed_tokens.as_linear(out)
+        out = out * self.model.config.logit_scale
+        return LanguageModelOutput(logits=out)
+
+    def make_cache(self):
+        caches = []
+        for i in range(self.config.num_hidden_layers):
+            if (
+                i % self.config.sliding_window_pattern
+                == self.config.sliding_window_pattern - 1
+            ):
+                caches.append(KVCache())
+            else:
+                caches.append(
+                    RotatingKVCache(max_size=self.config.sliding_window, keep=0)
+                )
+        return caches
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    @property
+    def head_dim(self):
+        return self.model.config.head_dim
+
+    @property
+    def n_kv_heads(self):
+        return self.model.config.num_key_value_heads