nexaai 1.0.19rc7__cp310-cp310-macosx_14_0_universal2.whl → 1.0.19rc8__cp310-cp310-macosx_14_0_universal2.whl

nexaai/binds/nexa_mlx/py-lib/vlm/modeling/models/llava_next/vision.py

@@ -1,243 +0,0 @@
-import inspect
-import math
-from dataclasses import dataclass
-from typing import Optional
-
-import mlx.core as mx
-import mlx.nn as nn
-import numpy as np
-
-
-@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 = 336
-    patch_size: int = 14
-    projection_dim: int = 768
-    vocab_size: int = 32000
-    num_channels: int = 3
-    layer_norm_eps: float = 1e-5
-
-    @classmethod
-    def from_dict(cls, params):
-        return cls(
-            **{
-                k: v
-                for k, v in params.items()
-                if k in inspect.signature(cls).parameters
-            }
-        )
-
-
-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 = 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"
-            )
-
-        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 = 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, queries, keys, values, mask=None):
-        queries = self.q_proj(queries)
-        keys = self.k_proj(keys)
-        values = self.v_proj(values)
-
-        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="fast")
-        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
-        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
-
-    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.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:
-        y = self.layer_norm1(x)
-        y = self.self_attn(y, y, y, mask)
-        x = x + y
-        y = self.layer_norm2(x)
-        y = self.mlp(y)
-        return x + y
-
-
-class Encoder(nn.Module):
-    def __init__(self, config: VisionConfig):
-        super().__init__()
-        self.layers = [EncoderLayer(config) for _ in range(config.num_hidden_layers)]
-
-
-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.class_embedding = mx.zeros((config.hidden_size,))
-
-        self.patch_embedding = nn.Conv2d(
-            in_channels=config.num_channels,
-            out_channels=self.embed_dim,
-            kernel_size=self.patch_size,
-            stride=self.patch_size,
-            bias=False,
-        )
-
-        self.num_patches = (self.image_size // self.patch_size) ** 2
-        self.num_positions = self.num_patches + 1
-        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
-
-    def __call__(self, x: mx.array) -> 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)
-        embed_dim = patch_embeddings.shape[-1]
-        cls_embeddings = mx.broadcast_to(
-            self.class_embedding, (batch_size, 1, embed_dim)
-        )
-        position_ids = mx.array(np.arange(self.num_positions)[None, :])
-
-        embeddings = mx.concatenate((cls_embeddings, patch_embeddings), axis=1)
-        embeddings += self.position_embedding(position_ids)
-        return embeddings
-
-
-class ClipVisionModel(nn.Module):
-    def __init__(self, config: VisionConfig):
-        super().__init__()
-        self.embeddings = VisionEmbeddings(config)
-        self.pre_layrnorm = nn.LayerNorm(config.hidden_size)
-        self.encoder = Encoder(config)
-        self.post_layernorm = nn.LayerNorm(config.hidden_size)
-
-    def __call__(
-        self,
-        x: mx.array,
-        output_hidden_states: Optional[bool] = None,
-    ) -> mx.array:
-        x = self.embeddings(x)
-        x = self.pre_layrnorm(x)
-
-        encoder_states = (x,) if output_hidden_states else None
-
-        for l in self.encoder.layers:
-            x = l(x, mask=None)
-            if output_hidden_states:
-                encoder_states = encoder_states + (x,)
-
-        pooler_output = self.post_layernorm(x[:, 0, :])
-        return pooler_output, x, encoder_states
-
-
-class VisionModel(nn.Module):
-    def __init__(self, config: VisionConfig):
-        super().__init__()
-
-        self.model_type = config.model_type
-        if self.model_type != "clip_vision_model":
-            raise ValueError(f"Unsupported model type: {self.model_type}")
-
-        self.vision_model = ClipVisionModel(config)
-
-    def __call__(
-        self, x: mx.array, output_hidden_states: Optional[bool] = None
-    ) -> mx.array:
-        return self.vision_model(x, 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

nexaai/binds/nexa_mlx/py-lib/vlm/modeling/models/mistral3/__init__.py

@@ -1,8 +0,0 @@
-from .mistral3 import (
-    LanguageModel,
-    Model,
-    ModelConfig,
-    TextConfig,
-    VisionConfig,
-    VisionModel,
-)

nexaai/binds/nexa_mlx/py-lib/vlm/modeling/models/mistral3/mistral3.py

@@ -1,283 +0,0 @@
-import glob
-import inspect
-import json
-from dataclasses import dataclass
-from pathlib import Path
-from typing import List, Optional, Tuple, Union
-
-import mlx.core as mx
-import mlx.nn as nn
-import numpy as np
-
-from ..pixtral import LanguageModel
-from ..pixtral import Model as PixtralModel
-from ..pixtral import TextConfig, VisionConfig, VisionModel
-
-
-@dataclass
-class ModelConfig:
-    text_config: TextConfig
-    vision_config: VisionConfig
-    model_type: str
-    ignore_index: int = -100
-    image_token_index: int = 10
-    vision_feature_select_strategy: str = "full"
-    vision_feature_layer: int = -1
-    vocab_size: int = 32000
-    spatial_merge_size: int = 2
-    multimodal_projector_bias: bool = False
-    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
-            }
-        )
-
-
-def _pair(x) -> Tuple[int, int]:
-    """Convert input to a pair of values."""
-    if isinstance(x, (list, tuple)):
-        return tuple(x)
-    return (x, x)
-
-
-def unfold(
-    input: mx.array,
-    kernel_size: Union[int, Tuple[int, int], List[int]],
-    dilation: Union[int, Tuple[int, int], List[int]] = 1,
-    padding: Union[int, Tuple[int, int], List[int]] = 0,
-    stride: Union[int, Tuple[int, int], List[int]] = 1,
-) -> mx.array:
-    """
-    Extract sliding local blocks from a batched input tensor (MLX implementation).
-
-    This is equivalent to PyTorch's nn.functional.unfold or im2col operation.
-
-    Args:
-        input: Input tensor of shape (B, C, H, W)
-        kernel_size: Size of the sliding blocks
-        dilation: Controls the spacing between kernel elements
-        padding: Controls the amount of implicit padding
-        stride: Controls the stride between blocks
-
-    Returns:
-        Unfolded tensor of shape (B, C*kernel_height*kernel_width, L)
-        where L is the number of blocks
-    """
-    # Convert to pairs
-    kernel_size = _pair(kernel_size)
-    dilation = _pair(dilation)
-    padding = _pair(padding)
-    stride = _pair(stride)
-
-    # Input shape
-    batch_size, channels, height, width = input.shape
-
-    # Add padding if needed
-    if padding[0] > 0 or padding[1] > 0:
-        padding_shape = (
-            (0, 0),
-            (0, 0),
-            (padding[0], padding[0]),
-            (padding[1], padding[1]),
-        )
-        input = mx.pad(input, padding_shape)
-
-    # Calculate output dimensions
-    height_out = (
-        height + 2 * padding[0] - dilation[0] * (kernel_size[0] - 1) - 1
-    ) // stride[0] + 1
-    width_out = (
-        width + 2 * padding[1] - dilation[1] * (kernel_size[1] - 1) - 1
-    ) // stride[1] + 1
-
-    # Initialize output arrays
-    blocks = []
-
-    # Extract blocks
-    for i in range(
-        0, height + 2 * padding[0] - kernel_size[0] * dilation[0] + 1, stride[0]
-    ):
-        for j in range(
-            0, width + 2 * padding[1] - kernel_size[1] * dilation[1] + 1, stride[1]
-        ):
-            # Extract the block for all channels
-            block = []
-            for di in range(kernel_size[0]):
-                for dj in range(kernel_size[1]):
-                    h_idx = i + di * dilation[0]
-                    w_idx = j + dj * dilation[1]
-                    # Get the block for all channels and add to our list
-                    block.append(input[:, :, h_idx, w_idx])
-
-            # Stack the channel-blocks
-            block = mx.stack(block, axis=1)  # Shape: (B, k*k, C)
-            block = mx.transpose(block, [0, 2, 1])  # Shape: (B, C, k*k)
-            blocks.append(block)
-
-    # Stack all blocks together
-    result = mx.stack(blocks, axis=-1)  # Shape: (B, C, k*k, L)
-
-    # Reshape to match PyTorch's unfold output format: (B, C*k*k, L)
-    result = mx.reshape(
-        result,
-        (
-            batch_size,
-            channels * kernel_size[0] * kernel_size[1],
-            height_out * width_out,
-        ),
-    )
-
-    return result
-
-
-class Mistral3PatchMerger(nn.Module):
-    """
-    Learned merging of spatial_merge_size ** 2 patches
-    """
-
-    def __init__(self, config: ModelConfig):
-        super().__init__()
-        self.config = config
-
-        hidden_size = config.vision_config.hidden_size
-        self.spatial_merge_size = config.spatial_merge_size
-        self.patch_size = self.config.vision_config.patch_size
-        self.merging_layer = nn.Linear(
-            hidden_size * self.spatial_merge_size**2, hidden_size, bias=False
-        )
-
-    def __call__(self, image_features: mx.array, image_sizes: mx.array) -> mx.array:
-
-        image_sizes = [
-            (image_size[0] // self.patch_size, image_size[1] // self.patch_size)
-            for image_size in image_sizes
-        ]
-
-        tokens_per_image = [h * w for h, w in image_sizes]
-        d = image_features.shape[-1]
-        image_features = image_features.astype(mx.bfloat16)
-        image_sizes = mx.array(image_sizes)
-
-        # Split the image features into chunks based on tokens_per_image
-        split_indices = []
-        current_index = 0
-        for tokens in tokens_per_image:
-            split_indices.append(current_index + tokens)
-            current_index += tokens
-
-        # Perform the split
-        chunks = mx.split(image_features, split_indices[:-1], axis=1)
-
-        permuted_tensor = []
-        for image_index, image_tokens in enumerate(chunks):
-
-            # Reshape image_tokens into a 2D grid
-            if image_tokens.shape[1] > 0:
-                h, w = image_sizes[image_index].tolist()
-
-                image_grid = image_tokens.reshape(h, w, d).transpose(2, 0, 1)[None, ...]
-
-                grid = unfold(
-                    image_grid,
-                    kernel_size=self.spatial_merge_size,
-                    stride=self.spatial_merge_size,
-                )
-                grid = grid.reshape(d * self.spatial_merge_size**2, -1).T
-                permuted_tensor.append(grid)
-
-        image_features = mx.concatenate(permuted_tensor, axis=0)
-        image_features = self.merging_layer(image_features)
-        return image_features[None, ...]
-
-
-class Mistral3MultiModalProjector(nn.Module):
-    def __init__(self, config: ModelConfig):
-        super().__init__()
-
-        self.norm = nn.RMSNorm(config.vision_config.hidden_size)
-        self.patch_merger = Mistral3PatchMerger(config)
-
-        num_feature_layers = (
-            1
-            if isinstance(config.vision_feature_layer, int)
-            else len(config.vision_feature_layer)
-        )
-        self.linear_1 = nn.Linear(
-            config.vision_config.hidden_size * num_feature_layers,
-            config.text_config.hidden_size,
-            bias=config.multimodal_projector_bias,
-        )
-        self.gelu = nn.GELU()
-        self.linear_2 = nn.Linear(
-            config.text_config.hidden_size,
-            config.text_config.hidden_size,
-            bias=config.multimodal_projector_bias,
-        )
-
-    def __call__(self, x: mx.array, image_sizes: mx.array) -> mx.array:
-        x = self.norm(x)
-
-        x = self.patch_merger(x, image_sizes)
-        x = self.linear_1(x)
-        x = self.gelu(x)
-        x = self.linear_2(x)
-        return x
-
-
-class Model(PixtralModel):
-    def __init__(self, config: ModelConfig):
-        super().__init__(config)
-        self.config = config
-
-        self.multi_modal_projector = Mistral3MultiModalProjector(config)
-
-    def get_input_embeddings(
-        self,
-        input_ids: Optional[mx.array] = None,
-        pixel_values: Optional[mx.array] = None,
-        **kwargs,
-    ):
-        image_sizes = kwargs.get("image_sizes", None)
-
-        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)
-
-        # Get the output hidden states from the vision model
-        if isinstance(pixel_values, list):
-            pixel_values = mx.concatenate(
-                [mx.array(pv)[None, ...] for pv in pixel_values], axis=0
-            )
-        if pixel_values.ndim == 3:
-            pixel_values = pixel_values[None, ...]
-
-        # Pass pixel_values as list of images, as each image is individually run through conv2d and position encoding
-        # Reference code from transformers: https://github.com/huggingface/transformers/blob/main/src/transformers/models/pixtral/modeling_pixtral.py#L479C9-L479C21
-        # and mistral_inference: https://github.com/mistralai/mistral-inference/blob/main/src/mistral_inference/vision_encoder.py#L85
-        *_, hidden_states = self.vision_tower(
-            pixel_values.transpose(0, 2, 3, 1),
-            output_hidden_states=True,
-        )
-        # Select the hidden states from the desired layer
-        selected_image_feature = hidden_states[self.vision_feature_layer]
-
-        # Pass image features through the multi-modal projector
-        image_features = self.multi_modal_projector(selected_image_feature, image_sizes)
-
-        # Insert special image tokens in the input_ids
-        final_inputs_embeds = self.merge_input_ids_with_image_features(
-            self.config.image_token_index, image_features, inputs_embeds, input_ids
-        )
-        return final_inputs_embeds
-
-    @property
-    def layers(self):
-        return self.language_model.model.layers

nexaai/binds/nexa_mlx/py-lib/vlm/modeling/models/mllama/__init__.py

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