nexaai 1.0.29__cp310-cp310-macosx_14_0_universal2.whl

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

@@ -0,0 +1,526 @@
+import copy
+import inspect
+from dataclasses import dataclass
+from functools import partial
+from math import sqrt
+from typing import Dict, Optional, Union
+
+import cv2
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+
+from .sam import SAMEncoder
+
+
+@dataclass
+class VisionConfig:
+    model_type: str
+    num_hidden_layers: int = 24
+    hidden_size: int = 1024
+    intermediate_size: int = 4096
+    num_attention_heads: int = 16
+    image_size: int = 384
+    patch_size: int = 16
+    num_channels: int = 3
+    layer_norm_eps: float = 1e-5
+    cls: str = None
+    params: dict = None
+
+    def __post_init__(self):
+        if "high_res_cfg" in self.params:
+            self.image_size = self.params["high_res_cfg"]["image_size"]
+
+    @classmethod
+    def from_dict(cls, params):
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )
+
+
+@dataclass
+class MLPConfig:
+    hidden_size: int
+    intermediate_size: int
+
+
+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 AttentionPoolLatent(nn.Module):
+    """Attention pooling w/ latent query"""
+
+    def __init__(
+        self,
+        in_features: int,
+        out_features: int = None,
+        embed_dim: int = None,
+        num_heads: int = 8,
+        mlp_ratio: float = 4.0,
+        qkv_bias: bool = True,
+        qk_norm: bool = False,
+        latent_len: int = 1,
+        latent_dim: int = None,
+        pos_embed: str = "",
+        pool_type: str = "token",
+        norm_layer: Optional[nn.Module] = None,
+        drop: float = 0.0,
+    ):
+        super().__init__()
+
+        embed_dim = embed_dim or in_features
+        out_features = out_features or in_features
+        assert embed_dim % num_heads == 0
+        self.num_heads = num_heads
+        self.head_dim = embed_dim // num_heads
+        self.scale = self.head_dim**-0.5
+        self.pool = pool_type
+
+        self.latent_dim = latent_dim or embed_dim
+        self.latent_len = latent_len
+        self.latent = mx.zeros((self.latent_len, embed_dim))[None, :]
+
+        self.q = nn.Linear(embed_dim, embed_dim, bias=qkv_bias)
+        self.kv = nn.Linear(embed_dim, embed_dim * 2, bias=qkv_bias)
+        self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
+        self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
+        self.proj = nn.Linear(embed_dim, embed_dim)
+        self.proj_drop = nn.Dropout(drop)
+
+        if pos_embed == "abs":
+            spatial_len = self.feat_size
+            self.pos_embed = mx.zeros((spatial_len, in_features))
+        else:
+            self.pos_embed = None
+
+        self.norm = nn.LayerNorm(out_features)
+        config = MLPConfig(
+            hidden_size=embed_dim, intermediate_size=int(embed_dim * mlp_ratio)
+        )
+        self.mlp = MLP(config)
+
+    def __call__(self, x: mx.array):
+        B, N, C = x.shape
+
+        if self.pos_embed is not None:
+            x = x + self.pos_embed.unsqueeze(0).to(x.dtype)
+
+        q_latent = mx.array(self.latent)
+
+        q = (
+            self.q(q_latent)
+            .reshape(B, self.latent_len, self.num_heads, self.head_dim)
+            .transpose(0, 2, 1, 3)
+        )
+
+        kv = (
+            self.kv(x)
+            .reshape(B, N, 2, self.num_heads, self.head_dim)
+            .transpose(2, 0, 3, 1, 4)
+        )
+        k, v = mx.split(kv, 2, axis=0)
+
+        q, k = self.q_norm(q), self.k_norm(k)
+
+        x = mx.fast.scaled_dot_product_attention(
+            q, k[0], v[0], scale=(1.0 / sqrt(q.shape[-1])), mask=None
+        )
+
+        x = x.transpose(0, 2, 1, 3).reshape(B, self.latent_len, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+
+        x = x + self.mlp(self.norm(x))
+
+        # optional pool if latent seq_len > 1 and pooled output is desired
+        if self.pool == "token":
+            x = x[:, 0]
+        elif self.pool == "avg":
+            x = x.mean(1)
+        return x
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dims: int,
+        num_heads: int,
+        qkv_bias: bool = False,
+    ):
+        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"
+            )
+
+        self.num_heads = num_heads = num_heads
+        head_dim = dims // num_heads
+        self.scale = head_dim**-0.5
+
+        self.qkv = nn.Linear(dims, dims * 3, bias=qkv_bias)
+        self.proj = nn.Linear(dims, dims, bias=True)
+
+    def __call__(self, x, mask=None):
+        qkv = self.qkv(x)
+        queries, keys, values = mx.split(qkv, 3, axis=-1)
+
+        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.proj(output)
+
+
+class FastGELUActivation(nn.Module):
+    """
+    Applies GELU approximation that is slower than QuickGELU but more accurate. See: https://github.com/hendrycks/GELUs
+    """
+
+    def __call__(self, input: mx.array) -> mx.array:
+        return (
+            0.5
+            * input
+            * (1.0 + mx.tanh(np.sqrt(2 / np.pi) * (input + 0.044715 * (input**3))))
+        ).astype(input.dtype)
+
+
+class MLP(nn.Module):
+    def __init__(self, config: Union[VisionConfig, Dict], bias: bool = True):
+        super().__init__()
+        self.activation_fn = FastGELUActivation()
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size, bias=bias)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size, bias=bias)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = self.activation_fn(self.fc1(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.attn = Attention(
+            config.hidden_size, config.num_attention_heads, qkv_bias=True
+        )
+        self.norm1 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = MLP(config)
+        self.norm2 = nn.LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    def __call__(self, x: mx.array, mask: Optional[mx.array] = None) -> mx.array:
+        y = self.norm1(x)
+        y = self.attn(y, mask)
+        x = x + y
+        y = self.norm2(x)
+        y = self.mlp(y)
+        return x + y
+
+
+class VisionEmbeddings(nn.Module):
+    def __init__(self, config: VisionConfig, norm_layer: bool = False):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = config.image_size
+        self.patch_size = config.patch_size
+
+        self.proj = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            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.norm = (
+            nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+            if norm_layer
+            else nn.Identity()
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        patch_embeddings = self.proj(x)
+        patch_embeddings = mx.flatten(patch_embeddings, start_axis=1, end_axis=2)
+        return self.norm(patch_embeddings)
+
+
+class SigLipVisionModel(nn.Module):
+    def __init__(
+        self,
+        config: VisionConfig,
+        ignore_head: bool,
+        pre_norm: bool = False,
+        no_embed_class: bool = True,
+    ):
+        super().__init__()
+        self.num_prefix_tokens = 1
+        self.no_embed_class = False
+        self.dynamic_img_size = False
+        self.ignore_head = ignore_head
+        self.cls_token = None
+        self.reg_token = None
+        self.patch_embed = VisionEmbeddings(config)
+        self.norm_pre = nn.LayerNorm(config.hidden_size) if pre_norm else nn.Identity()
+        self.blocks = [EncoderLayer(config) for _ in range(config.num_hidden_layers)]
+        self.norm = nn.LayerNorm(config.hidden_size)
+        num_patches = self.patch_embed.num_patches
+        embed_len = (
+            num_patches if no_embed_class else num_patches + self.num_prefix_tokens
+        )
+        self.pos_embed = mx.random.normal((embed_len, config.hidden_size))[None, :]
+
+        norm_layer = partial(nn.LayerNorm, eps=1e-5)
+        self.attn_pool = AttentionPoolLatent(
+            config.hidden_size,
+            num_heads=config.num_attention_heads,
+            norm_layer=norm_layer,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        output_hidden_states: Optional[bool] = None,
+    ) -> mx.array:
+        x = self.patch_embed(x)
+        x += self.pos_embed
+        x = self.norm_pre(x)
+
+        encoder_states = (x,) if output_hidden_states else None
+        for l in self.blocks:
+            x = l(x, mask=None)
+            if output_hidden_states:
+                encoder_states = encoder_states + (x,)
+
+        pooler_output = self.norm(x)
+
+        if not self.ignore_head:
+            pooler_output = self.attn_pool(pooler_output)
+        return pooler_output, x, encoder_states
+
+
+class HybridVisionModel(nn.Module):
+    def __init__(self, config: VisionConfig, resolution: str, ignore_head: bool = True):
+        super().__init__()
+
+        self.model_type = config.model_type
+        self.resolution = resolution
+        if self.model_type != "vision":
+            raise ValueError(f"Unsupported model type: {self.model_type}")
+
+        if resolution == "high":
+            self.vision_tower = SAMEncoder()
+        else:
+            self.vision_tower = SigLipVisionModel(config, ignore_head)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        if self.resolution == "high":
+            return self.vision_tower(x)
+        else:
+            return self.vision_tower(x)[0]
+
+
+# def resize_image(image, size, antialias=True):
+#     """
+#     Resize an image using scipy.ndimage.zoom with an option for bicubic interpolation.
+
+#     Args:
+#         image (numpy.ndarray): The input image array.
+#         size (tuple): The target size as (width, height).
+#         antialias (bool): True to use bicubic interpolation, False to use nearest neighbor.
+
+#     Returns:
+#         numpy.ndarray: The resized image array.
+#     """
+#     # Ensure the image is an array and remove singleton dimensions
+#     image = np.array(image[0])
+
+#     # Calculate zoom factors for the spatial dimensions
+#     # Note: size is expected as (width, height) but image.shape gives (height, width)
+#     current_height, current_width = image.shape[:2]
+#     width_factor = size[0] / current_width
+#     height_factor = size[1] / current_height
+#     zoom_factors = (height_factor, width_factor)  # Apply zoom to height and width
+
+#     # Choose the interpolation order: 3 for bicubic, 0 for nearest
+#     order = 3 if antialias else 0
+
+#     # Apply zoom to the image. Handle both grayscale and color images.
+#     if image.ndim == 2:  # Grayscale image
+#         resized_image = zoom(image, zoom_factors, order=order)
+#     elif image.ndim == 3:  # Color image
+#         # Apply zoom separately for each channel
+#         resized_channels = [
+#             zoom(image[:, :, i], zoom_factors, order=order)
+#             for i in range(image.shape[2])
+#         ]
+#         resized_image = np.stack(resized_channels, axis=2)
+
+#     return resized_image
+
+
+# TODO: Match the output of scipy.ndimage.zoom
+def resize_image(image, size, antialias=True):
+    """
+    Resize an image with OpenCV.
+
+    Args:
+        image (numpy.ndarray): The input image array.  Supports H × W or H × W × C.
+                               If you pass in a batch (N × H × W × C) just slice the
+                               element you want, e.g. image[0].
+        size  (tuple): Target size as (width, height) — exactly the same order that
+                       cv2.resize expects.
+        antialias (bool):
+            * True  → high‑quality interpolation (bicubic for upscaling, area for downscaling)
+            * False → nearest‑neighbor (fast, blocky)
+
+    Returns:
+        numpy.ndarray: The resized image array.
+    """
+    img = np.ascontiguousarray(np.asarray(image))
+    if img.ndim == 4 and img.shape[0] == 1:  # squeeze stray batch dim
+        img = img[0]
+    h0, w0 = img.shape[:2]
+
+    # --- work out dsize vs fx/fy ---------------------------------------------
+    dsize = None
+    fx = fy = 0.0
+    if isinstance(size, (int, float)):  # uniform scale
+        fx = fy = float(size)
+    elif isinstance(size, (tuple, list)) and len(size) == 2:
+        a, b = size
+        # Heuristic: treat "small" floats as scale factors
+        if all(isinstance(x, (int, float)) and x < 10 for x in (a, b)):
+            fx, fy = float(a), float(b)  # scale factors
+        else:
+            dsize = (int(a), int(b))  # absolute pixels
+    else:
+        raise ValueError("target must be scalar or a 2‑tuple")
+
+    # Guard against zeros after int‑casting
+    if dsize:
+        if dsize[0] <= 0 or dsize[1] <= 0:
+            raise ValueError(f"dsize became {dsize}")
+    else:
+        if fx <= 0 or fy <= 0:
+            raise ValueError(f"fx,fy became {(fx, fy)}")
+
+    # --- choose interpolation -------------------------------------------------
+    if antialias:
+        # Use Lanczos interpolation for potentially better detail preservation
+        interp = cv2.INTER_LANCZOS4
+    else:
+        interp = cv2.INTER_NEAREST
+
+    # --- call OpenCV ----------------------------------------------------------
+    return mx.array(cv2.resize(img, dsize=dsize, fx=fx, fy=fy, interpolation=interp))
+
+
+class VisionModel(nn.Module):
+    def __init__(self, config: VisionConfig, ignore_head: bool = True):
+        super().__init__()
+
+        self.model_type = config.model_type
+        self.config = config
+        if self.model_type != "vision":
+            raise ValueError(f"Unsupported model type: {self.model_type}")
+
+        if config.cls == "HybridVisionTower":
+            self.high_layer_norm = nn.LayerNorm(
+                config.params["high_res_cfg"]["output_dim"]
+            )
+            self.low_layer_norm = nn.LayerNorm(
+                config.params["low_res_cfg"]["output_dim"]
+            )
+
+            high_res_cfg = copy.deepcopy(config)
+            high_res_cfg.image_size = config.params["high_res_cfg"]["image_size"]
+            self.vision_tower_high = HybridVisionModel(
+                high_res_cfg, "high", ignore_head
+            )
+
+            low_res_cfg = copy.deepcopy(config)
+            low_res_cfg.image_size = config.params["low_res_cfg"]["image_size"]
+
+            self.vision_tower_low = HybridVisionModel(low_res_cfg, "low", ignore_head)
+            self.low_res_size = config.params["low_res_cfg"]["image_size"]
+            self.resize = lambda image: resize_image(
+                image, (self.low_res_size, self.low_res_size), antialias=True
+            )
+
+        else:
+            self.vision_tower = SigLipVisionModel(config, ignore_head)
+
+    def __call__(
+        self, x: mx.array, output_hidden_states: Optional[bool] = None
+    ) -> mx.array:
+        if self.config.cls == "HybridVisionTower":
+            high_images = x
+            low_images = mx.array(self.resize(np.array(x)))[None, :]
+
+            high_res = self.vision_tower_high(high_images)
+            low_res = self.vision_tower_low(low_images)
+
+            return (high_res, low_res)
+        else:
+            return self.vision_tower(x, output_hidden_states)
+
+    def sanitize(self, weights):
+        sanitized_weights = {}
+        weight_keys = {
+            "neck.0.weight",
+            "neck.2.weight",
+            "neck_hd.0.weight",
+            "neck_hd.2.weight",
+            "downsamples.0.weight",
+            "downsamples.1.weight",
+            "patch_embed.proj.weight",
+        }
+        for k, v in weights.items():
+            if "position_ids" in k:
+                # Remove unused position_ids
+                continue
+
+            elif ".".join(k.split(".")[-3:]) in weight_keys:
+                # 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

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

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