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

mlx_vlm/models/idefics2/language.py

@@ -0,0 +1,161 @@
+from typing import Optional
+
+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
+from .config import TextConfig
+
+
+class Attention(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+
+        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
+
+        head_dim = config.hidden_size // n_heads
+        self.scale = head_dim**-0.5
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=False)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=False)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        self.rope = nn.RoPE(
+            head_dim,
+            traditional=config.rope_traditional,
+            base=config.rope_theta,
+        )
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[KVCache] = None,
+    ) -> mx.array:
+        B, L, D = x.shape
+
+        queries, keys, values = self.q_proj(x), self.k_proj(x), self.v_proj(x)
+
+        # Prepare the queries, keys and values for the attention computation
+        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)
+
+        if cache is not None:
+            queries = self.rope(queries, offset=cache.offset)
+            keys = self.rope(keys, offset=cache.offset)
+            keys, values = cache.update_and_fetch(keys, values)
+        else:
+            queries = self.rope(queries)
+            keys = self.rope(keys)
+
+        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.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+        self.up_proj = nn.Linear(dim, hidden_dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        return self.down_proj(nn.silu(self.gate_proj(x)) * self.up_proj(x))
+
+
+class TransformerBlock(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.num_attention_heads = config.num_attention_heads
+        self.hidden_size = config.hidden_size
+        self.self_attn = Attention(config)
+        self.mlp = MLP(config.hidden_size, config.intermediate_size)
+        self.input_layernorm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.post_attention_layernorm = nn.RMSNorm(
+            config.hidden_size, eps=config.rms_norm_eps
+        )
+        self.config = config
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[KVCache] = None,
+    ) -> mx.array:
+        r = self.self_attn(self.input_layernorm(x), mask, cache)
+        h = x + r
+        r = self.mlp(self.post_attention_layernorm(h))
+        out = h + r
+        return out
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.model_type = config.model_type
+        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=config) for _ in range(config.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        inputs_embeds: Optional[mx.array] = None,
+        mask: Optional[mx.array] = None,
+        cache=None,
+        **kwargs,
+    ):
+        # for passing merged input embeddings
+        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:
+            mask = create_attention_mask(h, cache)
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        logits = self.lm_head(self.norm(h))
+        return LanguageModelOutput(logits=logits)
+
+    def sanitize(self, weights):
+        # Remove unused precomputed rotary freqs
+        return {
+            k: v for k, v in weights.items() if "self_attn.rotary_emb.inv_freq" not in k
+        }
+
+    @property
+    def layers(self):
+        return self.model.layers
+
+    @property
+    def head_dim(self):
+        return self.config.hidden_size // self.config.num_attention_heads
+
+    @property
+    def n_kv_heads(self):
+        return self.config.num_key_value_heads

mlx_vlm/models/idefics2/vision.py

@@ -0,0 +1,244 @@
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+
+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,
+    ):
+        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=True)
+        self.k_proj = nn.Linear(key_input_dims, dims, bias=True)
+        self.v_proj = nn.Linear(value_input_dims, value_dims, bias=True)
+        self.out_proj = nn.Linear(value_dims, value_output_dims, bias=True)
+
+    def __call__(self, x: mx.array, mask=None):
+        B, L, _ = x.shape
+        queries = self.q_proj(x)
+        keys = self.k_proj(x)
+        values = self.v_proj(x)
+
+        num_heads = self.num_heads
+
+        queries = queries.reshape(B, L, num_heads, -1).transpose(0, 2, 1, 3)
+        keys = keys.reshape(B, L, num_heads, -1).transpose(0, 2, 1, 3)
+        values = values.reshape(B, L, num_heads, -1).transpose(0, 2, 1, 3)
+        if mask is not None:
+            mask = mask[:, :, mask.shape[-2] :, :]
+
+        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)
+        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, 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)]
+
+    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)
+
+
+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(
+            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
+        self.num_positions = self.num_patches**2
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+
+    def __call__(self, x: mx.array, mask: Optional[mx.array] = None) -> mx.array:
+        B, H, W, C = x.shape
+        patch_embeddings = self.patch_embedding(x)
+        patch_embeddings = mx.flatten(patch_embeddings, start_axis=1, end_axis=2)
+        max_nb_patches_h, max_nb_patches_w = (
+            H // self.patch_size,
+            W // self.patch_size,
+        )
+        boundaries = np.linspace(
+            1 / self.num_patches, 1.0, self.num_patches, endpoint=False
+        )
+        position_ids = np.zeros((B, max_nb_patches_h * max_nb_patches_w), dtype=int)
+
+        for batch_idx, p_attn_mask in enumerate(mask):
+            p_attn_mask = np.array(p_attn_mask)
+            nb_patches_h = p_attn_mask[:, 0].sum()
+            nb_patches_w = p_attn_mask[0, :].sum()
+
+            fractional_coords_h = np.linspace(0, 1, nb_patches_h, endpoint=False)
+            fractional_coords_w = np.linspace(0, 1, nb_patches_w, endpoint=False)
+
+            bucket_coords_h = (
+                np.digitize(fractional_coords_h, boundaries, right=True) - 1
+            )
+            bucket_coords_w = (
+                np.digitize(fractional_coords_w, boundaries, right=True) - 1
+            )
+
+            pos_ids = (
+                bucket_coords_h[:, None] * self.num_patches + bucket_coords_w
+            ).flatten()
+            position_ids[batch_idx][p_attn_mask.reshape(-1)] = pos_ids
+
+        embeddings = patch_embeddings
+        embeddings += self.position_embedding(mx.array(position_ids))
+        return embeddings
+
+
+class VisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.config = config
+        self.model_type = config.model_type
+        if self.model_type not in ["idefics2", "idefics2_vision"]:
+            raise ValueError(f"Unsupported model type: {self.model_type}")
+        self.embeddings = VisionEmbeddings(config)
+        self.encoder = Encoder(config)
+        self.post_layernorm = nn.LayerNorm(config.hidden_size)
+
+    def __call__(
+        self,
+        x: mx.array,
+        patch_attention_mask: Optional[mx.array] = None,
+        output_hidden_states: Optional[bool] = None,
+    ) -> mx.array:
+
+        B, L, D, C = x.shape
+        if patch_attention_mask is None:
+            patch_size = self.config.patch_size
+            patch_attention_mask = mx.ones(
+                (
+                    B,
+                    L // patch_size,
+                    D // patch_size,
+                ),
+                dtype=mx.bool_,
+            )
+
+        x = self.embeddings(x, mask=patch_attention_mask)
+        encoder_outputs = self.encoder(x=x, output_hidden_states=output_hidden_states)
+
+        pooler_output = self.post_layernorm(encoder_outputs[0])
+
+        return pooler_output, x, encoder_outputs[-1]
+
+    def sanitize(self, weights):
+        sanitized_weights = {}
+        for k, v in weights.items():
+            if "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/idefics3/__init__.py

@@ -0,0 +1,4 @@
+from .config import ModelConfig, TextConfig, VisionConfig
+from .idefics3 import Model
+from .language import LanguageModel
+from .vision import VisionModel

mlx_vlm/models/idefics3/config.py

@@ -0,0 +1,54 @@
+from dataclasses import dataclass
+from typing import List, Optional
+
+from ..base import BaseModelConfig
+
+
+@dataclass
+class TextConfig(BaseModelConfig):
+    model_type: str
+    hidden_size: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    num_key_value_heads: int
+    rope_theta: float = 1000000.0
+    num_hidden_layers: int = 32
+    rope_traditional: bool = False
+    max_position_embeddings: int = 4096
+    tie_word_embeddings: bool = False
+
+    def __post_init__(self):
+        if self.num_key_value_heads is None:
+            self.num_key_value_heads = self.num_attention_heads
+
+
+@dataclass
+class VisionConfig(BaseModelConfig):
+    model_type: str
+    hidden_size: int
+    num_attention_heads: int
+    patch_size: int
+    num_hidden_layers: int = 12
+    intermediate_size: int = 3072
+    image_size: int = 224
+    num_channels: int = 3
+    layer_norm_eps: float = 1e-6
+
+
+@dataclass
+class ModelConfig(BaseModelConfig):
+    text_config: TextConfig
+    vision_config: VisionConfig
+    model_type: str
+    ignore_index: int = -100
+    vocab_size: int = 128259
+    scale_factor: int = 2
+    image_token_id: int = 49153
+    image_token_index: Optional[int] = None
+    eos_token_id: Optional[List[int]] = None
+
+    def __post_init__(self):
+        if self.image_token_index is None:
+            self.image_token_index = self.image_token_id

mlx_vlm/models/idefics3/idefics3.py

@@ -0,0 +1,221 @@
+import re
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+
+from ..base import InputEmbeddingsFeatures
+from .config import ModelConfig
+from .language import LanguageModel
+from .vision import VisionModel
+
+
+def masked_scatter(
+    final_embedding: mx.array,
+    image_mask_expanded: mx.array,
+    scaled_image_features: mx.array,
+):
+    # Reshape the tensors to 1D
+    final_embedding_shape = final_embedding.shape
+    scaled_image_features_flattened = mx.flatten(scaled_image_features)
+    final_embedding_flattened = mx.flatten(final_embedding)
+    image_mask_expanded_flattened = mx.flatten(image_mask_expanded)
+
+    # Scatter the scaled image features into the special image token positions
+    image_positions = mx.array(np.where(image_mask_expanded_flattened)[0], mx.uint32)
+    final_embedding_flattened[image_positions] = scaled_image_features_flattened
+
+    # Reshape back to the original shape
+    final_embedding = mx.reshape(final_embedding_flattened, final_embedding_shape)
+
+    return final_embedding
+
+
+class MLP(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        input_size = config.vision_config.hidden_size * (config.scale_factor**2)
+        output_size = config.text_config.hidden_size
+        self.proj = nn.Linear(input_size, output_size, bias=False)
+
+    def __call__(self, x):
+        return self.proj(x)
+
+
+class Idefics3Connector(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.scale_factor = config.scale_factor
+        self.modality_projection = MLP(config)
+
+    def pixel_shuffle(self, x, scale_factor=2):
+        bsz, seq, embed_dim = x.shape
+        height = width = int(seq**0.5)
+        x = x.reshape(bsz, height, width, embed_dim)
+        x = x.reshape(bsz, height, int(width / scale_factor), embed_dim * scale_factor)
+        x = x.transpose(0, 2, 1, 3)
+        x = x.reshape(
+            bsz,
+            int(width / scale_factor),
+            int(height / scale_factor),
+            embed_dim * (scale_factor**2),
+        )
+        x = x.transpose(0, 2, 1, 3)
+        x = x.reshape(bsz, int(seq / (scale_factor**2)), embed_dim * (scale_factor**2))
+        return x
+
+    def __call__(self, image_hidden_states):
+        image_hidden_states = self.pixel_shuffle(image_hidden_states, self.scale_factor)
+        image_hidden_states = self.modality_projection(image_hidden_states)
+        return image_hidden_states
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.model_type = config.model_type
+        self.config = config
+
+        self.vision_model = VisionModel(config.vision_config)
+        self.language_model = LanguageModel(config.text_config)
+        self.connector = Idefics3Connector(config)
+
+    def get_input_embeddings(
+        self,
+        input_ids: Optional[mx.array] = None,
+        pixel_values: Optional[mx.array] = None,
+        **kwargs,
+    ):
+        pixel_attention_mask = kwargs.get("pixel_attention_mask", None)
+
+        if pixel_values is None:
+            return InputEmbeddingsFeatures(
+                inputs_embeds=self.language_model.embed_tokens(input_ids)
+            )
+
+        inputs_embeds = self.language_model.embed_tokens(input_ids)
+
+        batch_size, num_images, num_channels, height, width = pixel_values.shape
+        pixel_values = pixel_values.reshape(
+            batch_size * num_images, num_channels, height, width
+        )
+
+        # Remove padding images - padding image are full 0.
+        nb_values_per_image = np.prod(pixel_values.shape[1:])
+        real_images_mask = (pixel_values == 0.0).sum(
+            axis=(-1, -2, -3)
+        ) != nb_values_per_image
+        real_images_inds = np.where(real_images_mask)[0].tolist()
+        pixel_values = pixel_values[real_images_inds, ...]
+
+        if pixel_attention_mask is None:
+            pixel_attention_mask = mx.ones(
+                (pixel_values.size(0), pixel_values.size(2), pixel_values.size(3)),
+                dtype=mx.bool,
+            )
+        else:
+            # Remove padding images from the mask
+            pixel_attention_mask = pixel_attention_mask.reshape(
+                batch_size * num_images, height, width
+            )
+            pixel_attention_mask = pixel_attention_mask[real_images_inds]
+
+        patch_size = self.config.vision_config.patch_size
+        batch_size, height, width = pixel_attention_mask.shape
+
+        # Calculate number of patches
+        patches_h = height // patch_size
+        patches_w = width // patch_size
+
+        # Reshape to extract patches
+        reshaped = pixel_attention_mask[
+            :, : patches_h * patch_size, : patches_w * patch_size
+        ]
+        reshaped = reshaped.reshape(
+            batch_size, patches_h, patch_size, patches_w, patch_size
+        )
+        reshaped = reshaped.transpose(
+            0, 1, 3, 2, 4
+        )  # (batch, patches_h, patches_w, patch_size, patch_size)
+
+        # Sum over patch dimensions and check if any pixels are active
+        patch_attention_mask = reshaped.sum(axis=(-1, -2)) > 0
+
+        pooler_output, *_ = self.vision_model(
+            pixel_values.transpose(0, 2, 3, 1),
+            patch_attention_mask=patch_attention_mask,
+            output_hidden_states=True,
+        )
+
+        image_features = pooler_output.astype(pixel_values.dtype)
+        image_features = self.connector(image_features)
+
+        final_inputs_embeds = self._prepare_inputs_for_multimodal(
+            image_features, inputs_embeds, input_ids
+        )
+        return InputEmbeddingsFeatures(inputs_embeds=final_inputs_embeds)
+
+    def _prepare_inputs_for_multimodal(self, image_features, inputs_embeds, input_ids):
+        special_image_mask = input_ids == self.config.image_token_index
+        n_image_tokens = special_image_mask.sum()
+        special_image_mask = special_image_mask[..., None]
+        special_image_mask = mx.broadcast_to(special_image_mask, inputs_embeds.shape)
+
+        n_image_features = image_features.shape[0]
+        n_image_mask_elements = special_image_mask.sum()
+        if n_image_mask_elements != image_features.size:
+            raise ValueError(
+                f"Image features and image tokens do not match: tokens: {n_image_tokens}, features {n_image_features}"
+            )
+
+        inputs_embeds = masked_scatter(
+            inputs_embeds, special_image_mask, image_features
+        )
+
+        return inputs_embeds
+
+    @property
+    def layers(self):
+        return self.language_model.layers
+
+    def __call__(
+        self,
+        input_ids: mx.array,
+        pixel_values: mx.array,
+        cache=None,
+        **kwargs,
+    ):
+        input_embeddings_features = self.get_input_embeddings(
+            input_ids, pixel_values, **kwargs
+        )
+        logits = self.language_model(
+            inputs=input_ids,
+            cache=cache,
+            inputs_embeds=input_embeddings_features.inputs_embeds,
+        )
+        return logits
+
+    def sanitize(self, weights):
+        weights = {
+            (
+                f"{k.split('.', 1)[1]}"
+                if re.match(r"^model\.", k)
+                else (f"language_model.{k}" if re.match(r"^lm_head\.", k) else k)
+            ): v
+            for k, v in weights.items()
+        }
+
+        weights = {
+            (
+                f"language_model.{k.split('.', 1)[1]}"
+                if re.match(
+                    r"^text_model\.",
+                    k,
+                )
+                else k
+            ): v
+            for k, v in weights.items()
+        }
+
+        return weights