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

mlx_vlm/models/paligemma/paligemma.py

@@ -0,0 +1,140 @@
+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 PaliGemmaMultiModalProjector(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.linear = nn.Linear(
+            config.vision_config.hidden_size,
+            config.vision_config.projection_dim,
+            bias=True,
+        )
+
+    def __call__(self, x: mx.array) -> mx.array:
+        output = self.linear(x)
+        return output
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.model_type = config.model_type
+        self.config = config
+
+        self.vision_tower = VisionModel(config.vision_config)
+        self.language_model = LanguageModel(config.text_config)
+        self.multi_modal_projector = PaliGemmaMultiModalProjector(config)
+
+    def get_input_embeddings(
+        self,
+        input_ids: Optional[mx.array] = None,
+        pixel_values: Optional[mx.array] = None,
+        mask: Optional[mx.array] = None,
+        **kwargs,
+    ):
+        if pixel_values is None:
+            return InputEmbeddingsFeatures(
+                inputs_embeds=self.language_model.model.embed_tokens(input_ids)
+            )
+
+        inputs_embeds = self.language_model.model.embed_tokens(input_ids)
+
+        hidden_state, _, _ = self.vision_tower(
+            pixel_values.transpose(0, 2, 3, 1).astype(inputs_embeds.dtype),
+            output_hidden_states=True,
+        )
+
+        image_features = hidden_state[None, :].astype(pixel_values.dtype)
+        image_features = self.multi_modal_projector(image_features)
+
+        final_inputs_embeds, final_attention_mask_4d = (
+            self._prepare_inputs_for_multimodal(
+                image_features, inputs_embeds, input_ids, mask
+            )
+        )
+        return InputEmbeddingsFeatures(
+            inputs_embeds=final_inputs_embeds, attention_mask_4d=final_attention_mask_4d
+        )
+
+    def _prepare_inputs_for_multimodal(
+        self, image_features, inputs_embeds, input_ids, attention_mask
+    ):
+        _, _, embed_dim = image_features.shape
+
+        batch_size, sequence_length = input_ids.shape
+        scaled_image_features = image_features / (self.config.hidden_size**0.5)
+        final_embedding = mx.zeros((batch_size, sequence_length, embed_dim))
+
+        text_mask = (input_ids != self.config.image_token_index) & (
+            input_ids != self.config.pad_token_id
+        )
+        image_mask = input_ids == self.config.image_token_index
+        pad_mask = input_ids == self.config.pad_token_id
+
+        # expand masks to match embedding dimension
+        text_mask_expanded = mx.expand_dims(text_mask, -1)
+        text_mask_expanded = mx.repeat(text_mask_expanded, embed_dim, axis=-1)
+        pad_mask_expanded = mx.expand_dims(pad_mask, -1)
+        pad_mask_expanded = mx.repeat(pad_mask_expanded, embed_dim, axis=-1)
+
+        # insert padding and text token embeddings
+        final_embedding = mx.where(text_mask_expanded, inputs_embeds, final_embedding)
+        final_embedding = mx.where(
+            pad_mask_expanded, mx.zeros_like(final_embedding), final_embedding
+        )
+        pad_size = final_embedding.shape[1] - scaled_image_features.shape[1]
+        scaled_image_features = mx.pad(
+            scaled_image_features, ((0, 0), (0, pad_size), (0, 0))
+        )
+        # insert image embeddings - the image mask is always less or equal to the sentence in length
+        image_mask_expanded = mx.expand_dims(image_mask, -1)
+        image_mask_expanded = mx.repeat(image_mask_expanded, embed_dim, axis=-1)
+        final_embedding = mx.where(
+            image_mask_expanded, scaled_image_features, final_embedding
+        )
+
+        final_embedding = mx.where(
+            pad_mask_expanded, mx.zeros_like(final_embedding), final_embedding
+        )
+
+        attention_mask_expanded_1 = mx.expand_dims(attention_mask, 1)
+        attention_mask_expanded_2 = mx.expand_dims(attention_mask, 2)
+        final_attention_mask_4d = attention_mask_expanded_1 * attention_mask_expanded_2
+        final_attention_mask_4d = final_attention_mask_4d
+        final_attention_mask_4d = mx.expand_dims(final_attention_mask_4d, 1)
+        final_embedding = mx.array(final_embedding)
+        return final_embedding, final_attention_mask_4d
+
+    @property
+    def layers(self):
+        return self.language_model.model.layers
+
+    def __call__(
+        self,
+        input_ids: mx.array,
+        pixel_values: mx.array,
+        mask: Optional[mx.array] = None,
+        cache: Optional[mx.array] = None,
+        **kwargs,
+    ):
+        input_embeddings_features = self.get_input_embeddings(
+            input_ids, pixel_values, mask
+        )
+        input_embeddings = input_embeddings_features.inputs_embeds
+        final_attention_mask_4d = input_embeddings_features.attention_mask_4d
+
+        logits = self.language_model(
+            inputs=input_ids,
+            cache=cache,
+            inputs_embeds=input_embeddings,
+            mask=final_attention_mask_4d,
+        )
+        return logits

mlx_vlm/models/paligemma/vision.py

@@ -0,0 +1,218 @@
+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 = True,
+    ):
+        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=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, x, mask=None):
+        queries = self.q_proj(x)
+        keys = self.k_proj(x)
+        values = self.v_proj(x)
+
+        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="precise")
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size, bias=True)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size, bias=True)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = self.fc1(x)
+        x = self.activation_fn(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:
+        r = self.self_attn(self.layer_norm1(x), mask)
+        h = x + r
+        r = self.mlp(self.layer_norm2(h))
+        return h + r
+
+
+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[0]
+
+        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) ** 2
+        self.num_positions = self.num_patches
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+
+    def __call__(self, x: mx.array) -> mx.array:
+        patch_embeddings = self.patch_embedding(x)
+        patch_embeddings = mx.flatten(patch_embeddings, start_axis=1, end_axis=2)
+        position_ids = mx.array(np.arange(self.num_positions)[None, :])
+        embeddings = patch_embeddings
+        embeddings += self.position_embedding(position_ids)
+        return embeddings
+
+
+class SigLipVisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.embeddings = VisionEmbeddings(config)
+        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)
+
+        encoder_outputs = self.encoder(
+            x=x, output_hidden_states=output_hidden_states, mask=None
+        )
+
+        pooler_output = self.post_layernorm(encoder_outputs[0])
+
+        return pooler_output, x, encoder_outputs[-1]
+
+
+class VisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.model_type = config.model_type
+        if self.model_type != "siglip_vision_model":
+            raise ValueError(f"Unsupported model type: {self.model_type}")
+
+        self.vision_model = SigLipVisionModel(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/phi3_v/__init__.py

@@ -0,0 +1,5 @@
+from .config import ModelConfig, TextConfig, VisionConfig
+from .language import LanguageModel
+from .phi3_v import Model
+from .processing_phi3_v import Phi3VImageProcessor, Phi3VProcessor
+from .vision import VisionModel

mlx_vlm/models/phi3_v/config.py

@@ -0,0 +1,55 @@
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Union
+
+from ..base import BaseModelConfig
+
+
+@dataclass
+class ModelConfig(BaseModelConfig):
+    text_config: "TextConfig" = field(default_factory=lambda: TextConfig())
+    vision_config: "VisionConfig" = field(default_factory=lambda: VisionConfig())
+    model_type: str = "phi3_v"
+    vocab_size: int = 32064
+
+    num_hidden_layers: int = 32
+    intermediate_size: int = 8192
+    num_attention_heads: int = 32
+    rms_norm_eps: float = 1e-5
+
+    ignore_index: int = -100
+    image_token_index: int = 257152
+    hidden_size: int = 2048
+    pad_token_id: int = 0
+
+    num_key_value_heads: int = None
+    rope_theta: float = 10000
+    rope_traditional: bool = False
+    partial_rotary_factor: float = 1.0
+    rope_scaling: Optional[Dict[str, Union[float, str]]] = None
+    max_position_embeddings: int = 131072
+    original_max_position_embeddings: int = 4096
+    eos_token_id: Optional[List[int]] = None
+
+
+@dataclass
+class TextConfig(BaseModelConfig):
+    max_position_embeddings: int = 4096
+
+
+@dataclass
+class VisionConfig(BaseModelConfig):
+    model_type: str = "phi3_v"
+    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
+    image_dim_out: int = (1024,)
+    model_name: str = "openai/clip-vit-large-patch14-336"
+    name: str = "clip_vision_model"
+    num_img_tokens: int = 144

mlx_vlm/models/phi3_v/language.py

@@ -0,0 +1,2 @@
+class LanguageModel:
+    pass

mlx_vlm/models/phi3_v/phi3_v.py

@@ -0,0 +1,239 @@
+from typing import Optional
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+from mlx_lm.models.rope_utils import SuScaledRoPE
+
+from ..base import InputEmbeddingsFeatures, LanguageModelOutput, create_attention_mask
+from ..cache import KVCache
+
+# Import processor to register it with AutoProcessor
+from . import processing_phi3_v  # noqa: F401
+from .config import ModelConfig, TextConfig
+from .vision import VisionModel
+
+
+class Attention(nn.Module):
+    def __init__(self, config: ModelConfig):
+        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.num_hidden_layers = config.num_hidden_layers
+
+        self.head_dim = head_dim = config.hidden_size // n_heads
+        self.scale = head_dim**-0.5
+
+        op_size = n_heads * head_dim + 2 * (n_kv_heads * head_dim)
+        self.qkv_proj = nn.Linear(dim, op_size, bias=False)
+        self.o_proj = nn.Linear(n_heads * head_dim, dim, bias=False)
+
+        rope_dim = int(head_dim * config.partial_rotary_factor)
+
+        # Check for Su-scaled RoPE by type or presence of short/long factors
+        rope_type = config.rope_scaling.get("type") if config.rope_scaling else None
+        has_su_factors = (
+            config.rope_scaling
+            and "short_factor" in config.rope_scaling
+            and "long_factor" in config.rope_scaling
+        )
+
+        if rope_type == "su" or has_su_factors:
+            self.rope = SuScaledRoPE(
+                rope_dim,
+                base=config.rope_theta,
+                max_position_embeddings=config.max_position_embeddings,
+                original_max_position_embeddings=config.original_max_position_embeddings,
+                short_factor=config.rope_scaling["short_factor"],
+                long_factor=config.rope_scaling["long_factor"],
+            )
+        else:
+            rope_scale = 1.0
+            if config.rope_scaling and rope_type == "linear":
+                rope_scale = 1 / config.rope_scaling["factor"]
+            self.rope = nn.RoPE(
+                rope_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, _ = x.shape
+
+        qkv = self.qkv_proj(x)
+        query_pos = self.n_heads * self.head_dim
+        queries, keys, values = mx.split(
+            qkv, [query_pos, query_pos + self.n_kv_heads * self.head_dim], axis=-1
+        )
+
+        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 = 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, dim, hidden_dim):
+        super().__init__()
+        self.gate_up_proj = nn.Linear(dim, 2 * hidden_dim, bias=False)
+        self.down_proj = nn.Linear(hidden_dim, dim, bias=False)
+
+    def __call__(self, x) -> mx.array:
+        x = self.gate_up_proj(x)
+        gate, x = mx.split(x, 2, axis=-1)
+        return self.down_proj(nn.silu(gate) * 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 Phi3V(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.config = config
+        self.vocab_size = config.vocab_size
+        self.num_hidden_layers = config.num_hidden_layers
+        self.embed_tokens = nn.Embedding(config.vocab_size, config.hidden_size)
+        self.vision_embed_tokens = VisionModel(config)
+        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,
+    ):
+        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[0])
+
+        for layer, c in zip(self.layers, cache):
+            h = layer(h, mask, c)
+
+        return self.norm(h)
+
+
+class Model(nn.Module):
+    def __init__(self, config: ModelConfig):
+        super().__init__()
+        self.model_type = config.model_type
+        self.model = Phi3V(config)
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+        self.config = config
+
+    def __call__(
+        self,
+        inputs: mx.array,
+        inputs_embeds: Optional[mx.array] = None,
+        pixel_values=None,
+        mask=None,
+        cache=None,
+        **kwargs,
+    ):
+        if inputs_embeds is None:
+            input_embeddings_features = self.get_input_embeddings(
+                inputs, pixel_values, **kwargs
+            )
+            inputs_embeds = input_embeddings_features.inputs_embeds
+
+        out = self.model(inputs, inputs_embeds, mask=mask, cache=cache)
+        logits = self.lm_head(out)
+
+        return LanguageModelOutput(logits=logits)
+
+    def get_input_embeddings(
+        self,
+        inputs: mx.array,
+        pixel_values: Optional[mx.array] = None,
+        **kwargs,
+    ):
+        image_sizes = kwargs.get("image_sizes", None) if kwargs else None
+
+        # Get text embeddings
+        inputs_embeds = self.model.embed_tokens(inputs)
+
+        # Find positions where inputs < 0 (image token positions)
+        inputs_list = inputs.tolist()
+        p = np.argwhere(np.array(inputs_list) < 0).tolist()
+
+        if pixel_values is not None:
+            inputs_embeds = self.model.vision_embed_tokens(
+                pixel_values, inputs_embeds, image_sizes, p
+            )
+
+        return InputEmbeddingsFeatures(inputs_embeds=inputs_embeds)
+
+    @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
+
+    @property
+    def language_model(self):
+        return self
+
+    @property
+    def vision_model(self):
+        return self.model.vision_embed_tokens