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

mlx_vlm/models/llava/config.py

@@ -0,0 +1,61 @@
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Union
+
+from ..base import BaseModelConfig
+
+
+@dataclass
+class TextConfig(BaseModelConfig):
+    model_type: str
+    hidden_size: int = 4096
+    num_hidden_layers: int = 32
+    intermediate_size: int = 11008
+    num_attention_heads: int = 32
+    rms_norm_eps: float = 1e-6
+    vocab_size: int = 32000
+    num_key_value_heads: int = None
+    rope_theta: float = 10000
+    rope_traditional: bool = False
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    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
+
+        if self.rope_scaling:
+            required_keys = {"factor", "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 self.rope_scaling["type"] != "linear":
+                raise ValueError("rope_scaling 'type' currently only supports 'linear'")
+
+
+@dataclass
+class VisionConfig(BaseModelConfig):
+    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
+
+
+@dataclass
+class ModelConfig(BaseModelConfig):
+    text_config: TextConfig
+    vision_config: VisionConfig
+    model_type: str
+    ignore_index: int = -100
+    image_token_index: int = 32000
+    vision_feature_select_strategy: str = "default"
+    vision_feature_layer: int = -2
+    vocab_size: int = 32000
+    eos_token_id: Optional[List[int]] = None

mlx_vlm/models/llava/language.py

@@ -0,0 +1,200 @@
+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
+
+        self.repeats = n_heads // n_kv_heads
+
+        head_dim = config.hidden_size // n_heads
+        self.scale = head_dim**-0.5
+
+        if config.model_type == "qwen2":
+            attention_bias = True
+        else:
+            attention_bias = False
+
+        self.q_proj = nn.Linear(dim, n_heads * head_dim, bias=attention_bias)
+        self.k_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.v_proj = nn.Linear(dim, n_kv_heads * head_dim, bias=attention_bias)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        rope_scale = (
+            1 / config.rope_scaling["factor"]
+            if config.rope_scaling is not None
+            and config.rope_scaling["type"] == "linear"
+            else 1
+        )
+        self.rope = nn.RoPE(
+            head_dim,
+            traditional=config.rope_traditional,
+            base=config.rope_theta,
+            scale=rope_scale,
+        )
+
+    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 Llama(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=config) for _ in range(config.num_hidden_layers)
+        ]
+        self.norm = nn.RMSNorm(config.hidden_size, eps=config.rms_norm_eps)
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        inputs_embeds: Optional[mx.array] = None,
+        mask: Optional[mx.array] = None,
+        cache=None,
+    ):
+        # 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)
+
+        return self.norm(h)
+
+
+class LanguageModel(nn.Module):
+    def __init__(self, config: TextConfig):
+        super().__init__()
+        self.config = config
+        self.model_type = config.model_type
+        if self.model_type not in ["llama", "qwen2"]:
+            raise ValueError(
+                f"Model type {self.model_type} not supported. Supported types: 'llama', 'qwen2'"
+            )
+        self.model = Llama(config)
+        if not config.tie_word_embeddings:
+            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,
+    ):
+        out = self.model(inputs, mask=mask, cache=cache, inputs_embeds=inputs_embeds)
+        if self.config.tie_word_embeddings:
+            out = self.model.embed_tokens.as_linear(out)
+        else:
+            out = self.lm_head(out)
+        return LanguageModelOutput(logits=out)
+
+    @staticmethod
+    def sanitize(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/llava/llava.py

@@ -0,0 +1,132 @@
+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
+
+
+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 ouptut hidden states from the vision model
+        *_, 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]
+
+        if isinstance(self.vision_feature_layer, int):
+            if self.vision_feature_select_strategy == "default":
+                selected_image_feature = selected_image_feature[:, 1:]
+
+        else:
+            hs_pool = [
+                hidden_states[layer_idx] for layer_idx in self.vision_feature_layer
+            ]
+            # For default; crop CLS from each hidden state in the hidden state pool
+            if self.vision_feature_select_strategy == "default":
+                hs_pool = [hs[:, 1:] for hs in hs_pool]
+            selected_image_feature = mx.concatenate(hs_pool, axis=-1)
+
+        # 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 InputEmbeddingsFeatures(inputs_embeds=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 == image_token_index)[1].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
+        )
+
+        # Pad image_positions to match the length of reshaped_image_hidden_states
+        num_positions_needed = len(image_positions)
+
+        if reshaped_image_hidden_states.shape[0] > num_positions_needed:
+            # TODO: Think about how to handle this case
+            raise ValueError(
+                "Llava model supports only one image per input. Please check your input_ids and pixel_values."
+            )
+
+        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_embeddings_features = self.get_input_embeddings(input_ids, pixel_values)
+        logits = self.language_model(
+            input_ids,
+            mask=mask,
+            cache=cache,
+            inputs_embeds=input_embeddings_features.inputs_embeds,
+        )
+        return logits

mlx_vlm/models/llava/vision.py

@@ -0,0 +1,233 @@
+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,
+        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
+
+        if config.model_type == "siglip_vision_model":
+            bias = True
+            self.class_embedding = None
+        else:
+            bias = False
+            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=bias,
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = (
+            self.num_patches + 1
+            if config.model_type == "clip_vision_model"
+            else self.num_patches
+        )
+        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)
+        if self.config.model_type == "siglip_vision_model":
+            embeddings = patch_embeddings
+        else:
+            embed_dim = patch_embeddings.shape[-1]
+            cls_embeddings = mx.broadcast_to(
+                self.class_embedding, (batch_size, 1, embed_dim)
+            )
+            embeddings = mx.concatenate((cls_embeddings, patch_embeddings), axis=1)
+
+        position_ids = mx.array(np.arange(self.num_positions)[None, :])
+
+        embeddings += self.position_embedding(position_ids)
+        return embeddings
+
+
+class ClipVisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.config = config
+        self.embeddings = VisionEmbeddings(config)
+        if self.config.model_type == "clip_vision_model":
+            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)
+        if self.config.model_type == "clip_vision_model":
+            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 not in ["clip_vision_model", "siglip_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

mlx_vlm/models/llava_bunny/__init__.py

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