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

mlx_vlm/models/pixtral/pixtral.py

@@ -0,0 +1,208 @@
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+from ..base import InputEmbeddingsFeatures
+from .config import ModelConfig
+from .language import LanguageModel
+from .vision import VisionModel
+
+
+class LlavaMultiModalProjector(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.linear_1 = nn.Linear(
+            config.vision_config.hidden_size, config.text_config.hidden_size, bias=True
+        )
+        self.gelu = nn.GELU()
+        self.linear_2 = nn.Linear(
+            config.text_config.hidden_size, config.text_config.hidden_size, bias=True
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = self.linear_1(x)
+        x = self.gelu(x)
+        x = self.linear_2(x)
+        return x
+
+
+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 = LlavaMultiModalProjector(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 InputEmbeddingsFeatures(
+                inputs_embeds=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)
+
+        # 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 InputEmbeddingsFeatures(inputs_embeds=final_inputs_embeds)
+
+    @staticmethod
+    def merge_input_ids_with_image_features(
+        image_token_index, image_features, inputs_embeds, input_ids
+    ):
+        """Merge image features into input embeddings at image token positions.
+
+        Args:
+            image_token_index: Token ID for image placeholder
+            image_features: Vision features from the projector [1, num_features, hidden_dim]
+            inputs_embeds: Input embeddings [batch_size, seq_len, hidden_dim]
+            input_ids: Input token IDs [batch_size, seq_len]
+
+        Returns:
+            Updated input embeddings with image features inserted
+        """
+        # Remove the extra batch dimension from image_features if present
+        if image_features.ndim == 3 and image_features.shape[0] == 1:
+            image_features = image_features.squeeze(0)  # [num_features, hidden_dim]
+
+        # Positions of <image> tokens in input_ids
+        image_positions = input_ids == image_token_index
+
+        # Get dimensions
+        batch_size, seq_len = input_ids.shape
+
+        # Process each batch item
+        batch_outputs = []
+        feature_start_idx = 0
+
+        for batch_idx in range(batch_size):
+            # Get mask for this batch
+            image_mask = image_positions[batch_idx]
+            num_positions = mx.sum(image_mask).item()
+
+            if num_positions > 0:
+                # Extract features for this batch
+                batch_features = image_features[
+                    feature_start_idx : feature_start_idx + num_positions
+                ]
+
+                # Validate we have the right number of features
+                if batch_features.shape[0] != num_positions:
+                    raise ValueError(
+                        f"Number of image token positions ({num_positions}) does not match "
+                        f"number of image features ({batch_features.shape[0]}) for batch {batch_idx}"
+                    )
+
+                # Create indices for gathering
+                cumsum = mx.cumsum(image_mask.astype(mx.int32))
+                feature_indices = mx.where(image_mask, cumsum - 1, 0)
+
+                # Gather features
+                gathered_features = batch_features[feature_indices]
+
+                # Combine with original embeddings
+                image_mask_expanded = mx.expand_dims(image_mask, axis=-1)
+                batch_output = mx.where(
+                    image_mask_expanded, gathered_features, inputs_embeds[batch_idx]
+                )
+
+                feature_start_idx += num_positions
+            else:
+                # No image tokens in this batch item
+                batch_output = inputs_embeds[batch_idx]
+
+            batch_outputs.append(batch_output)
+
+        # Stack all batch outputs
+        return mx.stack(batch_outputs, axis=0)
+
+    @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_embeddings_features = self.get_input_embeddings(
+            input_ids, pixel_values, **kwargs
+        )
+        logits = self.language_model(
+            input_ids,
+            cache=cache,
+            inputs_embeds=input_embeddings_features.inputs_embeds,
+        )
+        return logits
+
+    def sanitize(self, weights):
+        def transform_key(key):
+            if "vision_tower" in key and "vision_model" not in key:
+                if "transformer" in key:
+                    key = key.replace("vision_tower", "vision_tower.vision_model")
+                if "patch_conv" in key:
+                    key = key.replace("vision_tower", "vision_tower.vision_model")
+                if "ln_pre" in key:
+                    key = key.replace("vision_tower", "vision_tower.vision_model")
+
+            elif "vision_encoder" in key and "vision_tower" not in key:
+                if "transformer" in key:
+                    key = key.replace(
+                        "model.vision_encoder", "vision_tower.vision_model"
+                    )
+                if "patch_conv" in key:
+                    key = key.replace(
+                        "model.vision_encoder", "vision_tower.vision_model"
+                    )
+                if "ln_pre" in key:
+                    key = key.replace(
+                        "model.vision_encoder", "vision_tower.vision_model"
+                    )
+
+            elif "model.language_model" in key and "language_model.model" not in key:
+                key = key.replace("model.language_model", "language_model.model")
+
+            elif "lm_head" in key and "language_model" not in key:
+                key = key.replace("lm_head", "language_model.lm_head")
+
+            elif "model.vision_projection" in key:
+                key = key.replace("model.vision_projection", "multi_modal_projector")
+
+            return key
+
+        return {transform_key(k): v for k, v in weights.items()}

mlx_vlm/models/pixtral/vision.py

@@ -0,0 +1,293 @@
+from typing import List, Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+
+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
+
+
+def position_ids_in_meshgrid(patch_embeds_list, max_width):
+    positions = []
+    for patch in patch_embeds_list:
+        height, width = patch.shape[0], patch.shape[1]
+        h_grid, v_grid = mx.meshgrid(mx.arange(height), mx.arange(width), indexing="ij")
+        h_grid = h_grid.reshape(-1, 1)
+        v_grid = v_grid.reshape(-1, 1)
+        ids = h_grid * max_width + v_grid
+        positions.append(ids.flatten())
+    return mx.concatenate(positions)
+
+
+def generate_block_attention_mask(patch_embeds_list, tensor):
+    seq_len = tensor.shape[1]
+    d_min = -1e9  # Using a large negative value as MLX doesn't have finfo
+
+    causal_mask = mx.full((seq_len, seq_len), vals=d_min)
+
+    block_end_idx = mx.cumsum(mx.array(patch_embeds_list))
+    block_start_idx = mx.concatenate([mx.array([0]), mx.array(patch_embeds_list[:-1])])
+    block_start_idx = mx.cumsum(block_start_idx)
+
+    for start, end in zip(block_start_idx, block_end_idx):
+        start, end = int(start), int(end)  # Convert to integers for indexing
+        causal_mask[start:end, start:end] = 0
+
+    causal_mask = mx.broadcast_to(
+        causal_mask[None, None, :, :], (tensor.shape[0], 1, seq_len, seq_len)
+    )
+    return causal_mask.astype(tensor.dtype)
+
+
+def rotate_half(x):
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return mx.concatenate((-x2, x1), axis=-1)
+
+
+def apply_rotary_pos_emb(q, k, cos, sin, unsqueeze_dim=1):
+    cos = mx.expand_dims(cos, axis=unsqueeze_dim)
+    sin = mx.expand_dims(sin, axis=unsqueeze_dim)
+    q_embed = (q * cos) + (rotate_half(q) * sin)
+    k_embed = (k * cos) + (rotate_half(k) * sin)
+    return q_embed, k_embed
+
+
+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.embed_dim = dims
+        self.num_heads = num_heads
+        self.head_dim = self.embed_dim // self.num_heads
+
+        self.scale = self.head_dim**-0.5
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
+        self.o_proj = nn.Linear(self.embed_dim, self.embed_dim, bias=False)
+
+    def __call__(self, queries, keys, values, position_embeddings, 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)
+
+        cos, sin = position_embeddings
+        queries, keys = apply_rotary_pos_emb(queries, keys, cos, sin, unsqueeze_dim=0)
+
+        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.o_proj(output)
+
+
+class MLP(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        dim = config.hidden_size
+        hidden_dim = config.intermediate_size
+        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 EncoderLayer(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.attention = Attention(
+            config.hidden_size, config.num_attention_heads, bias=True
+        )
+        self.attention_norm = nn.RMSNorm(self.embed_dim, eps=config.rms_norm_eps)
+        self.feed_forward = MLP(config)
+        self.ffn_norm = nn.RMSNorm(self.embed_dim, eps=config.rms_norm_eps)
+
+    def __call__(
+        self,
+        x: mx.array,
+        position_embeddings: mx.array,
+        mask: Optional[mx.array] = None,
+    ) -> mx.array:
+        y = self.attention_norm(x)
+        y = self.attention(y, y, y, position_embeddings, mask)
+        x = x + y
+        y = self.ffn_norm(x)
+        y = self.feed_forward(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 PixtralRotaryEmbedding:
+    def __init__(self, config):
+        self.dim = config.head_dim
+        self.base = config.rope_theta
+        max_patches_per_side = config.image_size // config.patch_size
+        freqs = 1.0 / (
+            self.base ** (mx.arange(0, self.dim, 2).astype(mx.float32) / self.dim)
+        )
+
+        h = mx.arange(max_patches_per_side)
+        w = mx.arange(max_patches_per_side)
+
+        freqs_h = mx.outer(h, freqs[::2]).astype(mx.float32)
+        freqs_w = mx.outer(w, freqs[1::2]).astype(mx.float32)
+        inv_freq = mx.concatenate(
+            [
+                mx.tile(freqs_h[:, None, :], (1, max_patches_per_side, 1)),
+                mx.tile(freqs_w[None, :, :], (max_patches_per_side, 1, 1)),
+            ],
+            axis=-1,
+        ).reshape(-1, self.dim // 2)
+
+        self.inv_freq = mx.concatenate((inv_freq, inv_freq), axis=-1)
+
+    def __call__(self, x, position_ids):
+        freqs = self.inv_freq[position_ids]
+        emb = freqs
+        cos = mx.cos(emb)
+        sin = mx.sin(emb)
+        return cos.astype(x.dtype), sin.astype(x.dtype)
+
+
+class PixtralVisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.config = config
+        self.patch_conv = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=config.hidden_size,
+            kernel_size=config.patch_size,
+            stride=config.patch_size,
+            bias=False,
+        )
+        self.ln_pre = nn.RMSNorm(config.hidden_size)
+        self.transformer = Encoder(config)
+        self.patch_positional_embedding = PixtralRotaryEmbedding(config)
+
+    def __call__(
+        self,
+        x: List[mx.array],
+        output_hidden_states: Optional[bool] = None,
+    ) -> mx.array:
+
+        if x.dtype != self.patch_conv.weight.dtype:
+            x = x.astype(self.patch_conv.weight.dtype)
+
+        patch_embeds_list = self.patch_conv(x)
+        patch_embeds = patch_embeds_list.reshape(1, -1, patch_embeds_list.shape[-1])
+
+        patch_embeds = self.ln_pre(patch_embeds)
+
+        position_ids = position_ids_in_meshgrid(
+            patch_embeds_list,
+            max_width=self.config.image_size // self.config.patch_size,
+        )
+
+        position_embedding = self.patch_positional_embedding(patch_embeds, position_ids)
+
+        mask = generate_block_attention_mask(
+            [p.shape[1] * p.shape[0] for p in patch_embeds_list], patch_embeds
+        )
+
+        encoder_states = (patch_embeds,) if output_hidden_states else None
+
+        for l in self.transformer.layers:
+            patch_embeds = l(
+                patch_embeds, mask=mask, position_embeddings=position_embedding
+            )
+            if output_hidden_states:
+                encoder_states = encoder_states + (patch_embeds,)
+
+        return patch_embeds, encoder_states
+
+
+class VisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+
+        self.model_type = config.model_type
+        if self.model_type not in ["clip_vision_model", "pixtral"]:
+            raise ValueError(f"Unsupported model type: {self.model_type}")
+
+        self.vision_model = PixtralVisionModel(config)
+
+    def __call__(
+        self, x: List[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_conv.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/qwen2_5_vl/__init__.py

@@ -0,0 +1,2 @@
+from .config import ModelConfig, TextConfig, VisionConfig
+from .qwen2_5_vl import LanguageModel, Model, VisionModel

mlx_vlm/models/qwen2_5_vl/config.py

@@ -0,0 +1,90 @@
+import inspect
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Union
+
+from ..base import BaseModelConfig
+
+
+@dataclass
+class VisionConfig(BaseModelConfig):
+    model_type: str = "qwen2_5_vl"
+    depth: int = 32
+    hidden_size: int = 1280
+    intermediate_size: int = 3420
+    out_hidden_size: int = 1536
+    num_heads: int = 16
+    image_size: int = 384
+    patch_size: int = 14
+    vocab_size: int = 32000
+    mlp_ratio: float = 4.0
+    in_channels: int = 3
+    layer_norm_eps: float = 1e-6
+    spatial_patch_size: int = 14
+    spatial_merge_size: int = 2
+    tokens_per_second: int = 2
+    temporal_patch_size: int = 2
+    window_size: int = 112
+    patch_size: int = 14
+    fullatt_block_indexes: list[int] = field(default_factory=lambda: [7, 15, 23, 31])
+
+
+@dataclass
+class TextConfig(BaseModelConfig):
+    model_type: str
+    hidden_size: int
+    num_hidden_layers: int
+    intermediate_size: int
+    num_attention_heads: int
+    rms_norm_eps: float
+    vocab_size: int
+    num_key_value_heads: Optional[int] = None
+    max_position_embeddings: Optional[int] = 128000
+    rope_theta: float = 1000000.0
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    tie_word_embeddings: bool = True
+
+    def __post_init__(self):
+        if self.num_key_value_heads is None:
+            self.num_key_value_heads = self.num_attention_heads
+
+        if self.rope_scaling:
+            required_keys = {"mrope_section", "type"}
+            if not all(key in self.rope_scaling for key in required_keys):
+                raise ValueError(f"rope_scaling must contain keys {required_keys}")
+
+            if not self.rope_scaling["type"] in ["mrope", "default"]:
+                raise ValueError(f"rope_scaling type must be 'mrope' or 'default'")
+
+
+@dataclass
+class ModelConfig(BaseModelConfig):
+    text_config: TextConfig
+    vision_config: VisionConfig
+    model_type: str
+    ignore_index: int = -100
+    image_token_id: int = 151655
+    video_token_id: int = 151656
+    vision_start_token_id: int = 151652
+    vision_end_token_id: int = 151653
+    vision_token_id: int = 151654
+    vision_feature_select_strategy: str = "default"
+    vision_feature_layer: int = -2
+    vocab_size: int = 32000
+    eos_token_id: Optional[List[int]] = None
+
+    @classmethod
+    def from_dict(cls, params):
+        # Copy text config parameters from root level
+        excluded_keys = {"vision_config"}
+        params["text_config"] = dict(
+            filter(lambda x: x[0] not in excluded_keys, params.items())
+        )
+
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )