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

mlx_vlm/models/moondream2/vision.py

@@ -0,0 +1,200 @@
+from typing import Optional, Tuple
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+from PIL import Image
+
+from .config import VisionConfig
+
+
+def create_patches(pixel_values: mx.array, patch_size: int) -> mx.array:
+    """
+    Convert [B, C, H, W] images to [B, num_patches, patch_dim] patch sequences.
+    
+    This matches the PyTorch reference implementation exactly.
+    """
+    B, C, H, W = pixel_values.shape
+    P = patch_size
+    
+    # Reshape to [B, C, H/P, P, W/P, P]
+    x = pixel_values.reshape(B, C, H // P, P, W // P, P)
+    
+    # Permute to [B, H/P, W/P, C, P, P]
+    x = x.transpose(0, 2, 4, 1, 3, 5)
+    
+    # Flatten to [B, (H/P)*(W/P), C*P*P]
+    num_patches = (H // P) * (W // P)
+    patch_dim = C * P * P
+    x = x.reshape(B, num_patches, patch_dim)
+    
+    return x
+
+
+class Attention(nn.Module):
+    """Multi-head attention with combined QKV projection."""
+
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.hidden_size = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.hidden_size // self.num_heads
+        self.scale = self.head_dim**-0.5
+
+        # Combined QKV projection (like original moondream)
+        self.qkv = nn.Linear(self.hidden_size, 3 * self.hidden_size, bias=True)
+        self.proj = nn.Linear(self.hidden_size, self.hidden_size, bias=True)
+
+    def __call__(self, x: mx.array, mask: Optional[mx.array] = None) -> mx.array:
+        B, L, _ = x.shape
+
+        # Combined QKV projection then split
+        qkv = self.qkv(x)
+        q, k, v = mx.split(qkv, 3, axis=-1)
+
+        # Reshape for multi-head attention
+        q = q.reshape(B, L, self.num_heads, self.head_dim).transpose(0, 2, 1, 3)
+        k = k.reshape(B, L, self.num_heads, self.head_dim).transpose(0, 2, 1, 3)
+        v = v.reshape(B, L, self.num_heads, self.head_dim).transpose(0, 2, 1, 3)
+
+        # Scaled dot-product attention
+        output = mx.fast.scaled_dot_product_attention(
+            q, k, v, scale=self.scale, mask=mask
+        )
+        output = output.transpose(0, 2, 1, 3).reshape(B, L, -1)
+
+        return self.proj(output)
+
+
+class MLP(nn.Module):
+    """Feed-forward network with GELU activation."""
+
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size, bias=True)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size, bias=True)
+        self.activation = nn.GELU(approx="precise")
+
+    def __call__(self, x: mx.array) -> mx.array:
+        x = self.activation(self.fc1(x))
+        x = self.fc2(x)
+        return x
+
+
+class EncoderLayer(nn.Module):
+    """
+    Single transformer encoder layer with POST-NORM architecture.
+    
+    Key difference from standard: residual addition happens BEFORE normalization.
+    Pattern: x = x + attn(ln(x))
+    """
+
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.ln1 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.attn = Attention(config)
+        self.ln2 = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+        self.mlp = MLP(config)
+
+    def __call__(self, x: mx.array, mask: Optional[mx.array] = None) -> mx.array:
+        # Post-norm: x = x + sublayer(norm(x))
+        x = x + self.attn(self.ln1(x), mask)
+        x = x + self.mlp(self.ln2(x))
+        return x
+
+
+class VisionEncoder(nn.Module):
+    """Vision encoder with transformer layers."""
+
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.layers = [EncoderLayer(config) for _ in range(config.num_hidden_layers)]
+
+    def __call__(
+        self,
+        x: mx.array,
+        mask: Optional[mx.array] = None,
+        output_hidden_states: bool = False,
+    ):
+        encoder_states = (x,) if output_hidden_states else None
+
+        for layer in self.layers:
+            x = layer(x, mask)
+            if output_hidden_states:
+                encoder_states = encoder_states + (x,)
+
+        return x, encoder_states
+
+
+class VisionModel(nn.Module):
+    """
+    Moondream2 vision encoder.
+    
+    Architecture:
+    1. Linear patch embedding (not Conv2d): 588 (14×14×3) -> 1152
+    2. Add learnable positional embeddings
+    3. 27 transformer layers with post-norm
+    4. Final layer norm
+    
+    Reference: moondream2/vision.py (PyTorch)
+    """
+
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.config = config
+        
+        # Patch embedding: linear projection of flattened patches
+        patch_dim = config.patch_size * config.patch_size * config.num_channels  # 588
+        self.patch_emb = nn.Linear(patch_dim, config.hidden_size, bias=True)
+        
+        # Transformer encoder
+        self.encoder = VisionEncoder(config)
+        
+        # Post layer norm
+        self.post_layernorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+        # Positional embedding: [1, num_patches, hidden_size]
+        # This will be loaded from weights
+        num_patches = (config.image_size // config.patch_size) ** 2  # 729
+        self.position_embedding = mx.zeros((1, num_patches, config.hidden_size))
+
+    def __call__(
+        self,
+        pixel_values: mx.array,
+        output_hidden_states: bool = False,
+    ) -> mx.array:
+        """
+        Args:
+            pixel_values: [B, C, H, W] input images (normalized to [-1, 1])
+        Returns:
+            [B, num_patches, hidden_size] vision features
+        """
+        # Create patches: [B, C, H, W] -> [B, num_patches, patch_dim]
+        x = create_patches(pixel_values, self.config.patch_size)
+        
+        # Linear projection: [B, num_patches, patch_dim] -> [B, num_patches, hidden_size]
+        x = self.patch_emb(x)
+
+        # Add positional embedding
+        x = x + self.position_embedding
+
+        # Encode through transformer layers
+        x, encoder_states = self.encoder(x, output_hidden_states=output_hidden_states)
+
+        # Final layer norm
+        x = self.post_layernorm(x)
+
+        if output_hidden_states:
+            return x, encoder_states
+        return x
+
+    def sanitize(self, weights):
+        """Sanitize vision encoder weights."""
+        sanitized_weights = {}
+        for k, v in weights.items():
+            if "position_ids" in k:
+                # Skip position_ids
+                continue
+            else:
+                sanitized_weights[k] = v
+        return sanitized_weights

mlx_vlm/models/multi_modality/__init__.py

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

mlx_vlm/models/multi_modality/config.py

@@ -0,0 +1,108 @@
+import inspect
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple, Union
+
+from ..base import BaseModelConfig
+
+
+@dataclass
+class ProjectorConfig(BaseModelConfig):
+    cls: str
+    model_type: str
+    params: dict
+
+
+@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 = 102400
+    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
+
+    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 = 384
+    patch_size: int = 16
+    num_channels: int = 3
+    layer_norm_eps: float = 1e-5
+    cls: str = None
+    params: dict = None
+
+    def __post_init__(self):
+        if "high_res_cfg" in self.params:
+            self.image_size = self.params["high_res_cfg"]["image_size"]
+
+
+@dataclass
+class MLPConfig(BaseModelConfig):
+    hidden_size: int
+    intermediate_size: int
+
+
+@dataclass
+class SAMViTCfg:
+    image_size: Union[Tuple[int, int], int] = 1024
+    width: int = 768
+    layers: int = 12
+    heads: int = 12
+    patch_size: int = 16
+    window_size: int = 14
+    prompt_embed_dim: int = 256
+    global_attn_indexes: Union[List[int], Tuple[int]] = (2, 5, 8, 11)
+    downsample_channels: Union[List[int], Tuple[int]] = (512, 1024)
+
+
+@dataclass
+class ModelConfig(BaseModelConfig):
+    text_config: TextConfig
+    vision_config: VisionConfig
+    projector_config: ProjectorConfig
+    model_type: str
+    ignore_index: int = -100
+    image_token_index: int = 100015
+    vision_feature_select_strategy: str = "default"
+    select_layer: int = -1
+    pad_id: int = 100001
+    num_image_tokens: int = 576
+    vocab_size: int = 32000
+    eos_token_id: Optional[List[int]] = None
+
+    @classmethod
+    def from_dict(cls, params):
+        if "aligner_config" in params:
+            params["projector_config"] = params["aligner_config"]
+            del params["aligner_config"]
+
+        return cls(
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+            }
+        )

mlx_vlm/models/multi_modality/language.py

@@ -0,0 +1,191 @@
+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
+
+        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)
+
+        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 != "llama":
+            raise ValueError(
+                f"Model type {self.model_type} not supported. Currently only 'llama' is supported"
+            )
+        self.model = Llama(config)
+        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)
+        logits = self.lm_head(out)
+        return LanguageModelOutput(logits=logits)
+
+    @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