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

mlx_vlm/models/deepseekocr/vision.py

@@ -0,0 +1,263 @@
+import math
+from typing import Dict, Optional, Union
+
+import mlx.core as mx
+import mlx.nn as nn
+import numpy as np
+
+from ..base import interpolate
+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,
+        qkv_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"
+            )
+
+        self.num_heads = num_heads = num_heads
+        head_dim = dims // num_heads
+        self.scale = head_dim**-0.5
+
+        self.qkv_proj = nn.Linear(dims, dims * 3, bias=qkv_bias)
+        self.out_proj = nn.Linear(dims, dims, bias=True)
+
+    def __call__(self, x, mask=None):
+        qkv = self.qkv_proj(x)
+        queries, keys, values = mx.split(qkv, 3, axis=-1)
+
+        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: Union[VisionConfig, Dict], bias: bool = True):
+        super().__init__()
+        self.activation_fn = nn.GELU()
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size, bias=bias)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size, bias=bias)
+
+    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, qkv_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, mask)
+        x = x + y
+        y = self.layer_norm2(x)
+        y = self.mlp(y)
+        return x + y
+
+
+class VisionEmbeddings(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.image_size = 224
+        self.patch_size = config.patch_size
+
+        self.class_embedding = mx.random.normal((self.embed_dim,))
+
+        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=False,
+        )
+
+        self.num_patches = (self.image_size // self.patch_size) ** 2
+        self.num_positions = self.num_patches + 1
+        self.position_embedding = nn.Embedding(self.num_positions, self.embed_dim)
+
+    def _get_abs_pos(self, abs_pos, tgt_size):
+        """
+        Resize absolute positional embeddings
+
+        Args:
+            abs_pos: Tensor of shape (L, C) - absolute position embeddings
+            tgt_size: int - target size M
+
+        Returns:
+            Tensor of shape (M, C) - resized position embeddings
+        """
+        dim = abs_pos.shape[-1]
+        abs_pos_new = mx.squeeze(abs_pos, axis=0)
+        cls_token, old_pos_embed = abs_pos_new[:1], abs_pos_new[1:]
+        src_size = int(math.sqrt(abs_pos_new.shape[0] - 1))
+        tgt_size_2d = int(math.sqrt(tgt_size))
+        dtype = abs_pos.dtype
+
+        if src_size != tgt_size_2d:
+            # Reshape to (1, src_size, src_size, dim) then transpose to (1, dim, src_size, src_size)
+            old_pos_embed = mx.reshape(old_pos_embed, (1, src_size, src_size, dim))
+            old_pos_embed = mx.transpose(old_pos_embed, (0, 3, 1, 2))
+            old_pos_embed = old_pos_embed.astype(mx.float32)
+
+            new_pos_embed = interpolate(old_pos_embed, (tgt_size_2d, tgt_size_2d))
+
+            new_pos_embed = new_pos_embed.astype(dtype)
+            new_pos_embed = mx.transpose(new_pos_embed, (0, 2, 3, 1))
+            new_pos_embed = mx.reshape(new_pos_embed, (tgt_size_2d * tgt_size_2d, dim))
+            vision_pos_embed = mx.concatenate([cls_token, new_pos_embed], axis=0)
+            vision_pos_embed = mx.reshape(
+                vision_pos_embed, (1, tgt_size_2d * tgt_size_2d + 1, dim)
+            )
+            return vision_pos_embed
+        else:
+            return abs_pos
+
+    def __call__(
+        self, x: mx.array, patch_embeds: Optional[mx.array] = None
+    ) -> mx.array:
+        batch_size, height, width, _ = x.shape
+        target_dtype = self.position_embedding.weight.dtype
+
+        if patch_embeds is not None:
+            patch_embeddings = patch_embeds
+        else:
+            patch_embeddings = self.patch_embedding(x)
+
+        # Flatten patch embeddings properly
+        patch_embeds = mx.flatten(patch_embeddings, start_axis=1, end_axis=2)
+
+        # Broadcast class embedding
+        class_embeds = mx.broadcast_to(
+            self.class_embedding, (batch_size, 1, self.embed_dim)
+        ).astype(target_dtype)
+
+        # Concatenate class and patch embeddings
+        embeddings = mx.concatenate([class_embeds, patch_embeds], axis=1)
+
+        # Create position IDs
+        position_ids = mx.array(np.arange(self.num_positions)[None, :])
+
+        # Add positional embeddings
+        embeddings = embeddings + self._get_abs_pos(
+            self.position_embedding(position_ids), embeddings.shape[1]
+        ).astype(target_dtype)
+
+        return embeddings
+
+
+class NoTPTransformer(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+        self.num_layers = config.layers
+        self.layers = [EncoderLayer(config) for _ in range(config.layers)]
+
+    def __call__(
+        self,
+        x: mx.array,
+    ) -> mx.array:
+        for l in self.layers:
+            x = l(x, mask=None)
+        return x
+
+
+class VisionModel(nn.Module):
+    def __init__(self, config: VisionConfig):
+        super().__init__()
+
+        self.model_type = config.model_type
+        self.config = config
+        if self.model_type != "vision":
+            raise ValueError(f"Unsupported model type: {self.model_type}")
+
+        self.embeddings = VisionEmbeddings(config)
+        self.pre_layrnorm = nn.LayerNorm(config.hidden_size)
+        self.transformer = NoTPTransformer(config)
+
+    def __call__(
+        self,
+        x: mx.array,
+        patch_embeds: mx.array = None,
+    ) -> mx.array:
+        x = self.embeddings(x, patch_embeds)
+        x = self.pre_layrnorm(x)
+        return self.transformer(x)
+
+    def sanitize(self, weights):
+        sanitized_weights = {}
+        weight_keys = {
+            "neck.0.weight",
+            "neck.2.weight",
+            "neck_hd.0.weight",
+            "neck_hd.2.weight",
+            "sam_model.net_2.weight",
+            "sam_model.net_3.weight",
+            "downsamples.0.weight",
+            "downsamples.1.weight",
+            "patch_embed.proj.weight",
+            "embeddings.patch_embedding.weight",
+        }
+        for k, v in weights.items():
+            if "position_ids" in k:
+                # Remove unused position_ids
+                continue
+
+            elif ".".join(k.split(".")[-3:]) in weight_keys:
+                # 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/deepseekocr_2/__init__.py

@@ -0,0 +1,12 @@
+# Import shared LanguageModel from deepseekocr
+from ..deepseekocr.language import LanguageModel
+from .config import (
+    MLPConfig,
+    ModelConfig,
+    ProjectorConfig,
+    Qwen2EncoderConfig,
+    TextConfig,
+    VisionConfig,
+)
+from .deepseekocr_2 import DeepseekOCR2Processor, Model
+from .vision import Qwen2Decoder2Encoder, VisionModel

mlx_vlm/models/deepseekocr_2/config.py

@@ -0,0 +1,216 @@
+import inspect
+from dataclasses import dataclass
+from typing import Dict, List, Optional, Tuple, Union
+
+from ..base import BaseModelConfig
+
+
+@dataclass
+class TextConfig(BaseModelConfig):
+    model_type: str = "deepseek_v2"
+    vocab_size: int = 102400
+    hidden_size: int = 1280
+    intermediate_size: int = 6848
+    moe_intermediate_size: int = 896
+    num_hidden_layers: int = 30
+    num_attention_heads: int = 32
+    num_key_value_heads: int = 32
+    n_shared_experts: Optional[int] = 2
+    n_routed_experts: Optional[int] = 64
+    routed_scaling_factor: float = 1.0
+    kv_lora_rank: int = 512
+    q_lora_rank: int = 1536
+    qk_rope_head_dim: int = 0
+    v_head_dim: int = 128
+    qk_nope_head_dim: int = 0
+    topk_method: str = "greedy"
+    n_group: Optional[int] = 1
+    topk_group: Optional[int] = 1
+    num_experts_per_tok: Optional[int] = 6
+    moe_layer_freq: int = 1
+    first_k_dense_replace: int = 0
+    max_position_embeddings: int = 2048
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 10000.0
+    rope_traditional: bool = False
+    rope_scaling: Dict = None
+    attention_bias: bool = False
+    scoring_func: str = "softmax"
+    attn_type: str = "DeepseekV2Attention"
+
+    def __post_init__(self):
+        if self.qk_nope_head_dim == 0:
+            self.attn_type = "LlamaAttention"
+
+        if self.num_key_value_heads is None:
+            self.num_key_value_heads = self.num_attention_heads
+
+
+@dataclass
+class Qwen2EncoderConfig(BaseModelConfig):
+    """Configuration for the Qwen2 decoder-as-encoder in the vision model."""
+
+    dim: int = 896
+    layers: int = 24
+    heads: int = 14
+    kv_heads: int = 2
+    intermediate_size: int = 4864
+    rms_norm_eps: float = 1e-6
+    rope_theta: float = 1000000.0
+
+
+@dataclass
+class VisionConfig(BaseModelConfig):
+    model_type: str
+    layers: int = 24
+    width: int = 1152
+    hidden_size: int = 896
+    intermediate_size: int = 4096
+    num_attention_heads: int = 16
+    image_size: int = 1024
+    patch_size: int = 14
+    num_channels: int = 3
+    layer_norm_eps: float = 1e-6
+    mlp_ratio: float = 3.7362
+    cls: str = None
+    params: dict = None
+
+    @classmethod
+    def from_dict(cls, params):
+        # Parse width configuration for SAM and Qwen2
+        width = params.get("width", {})
+        qwen2_config = width.get("qwen2-0-5b", {})
+        sam_config = width.get("sam_vit_b", {})
+
+        # Build qwen2 params for VisionModel
+        qwen2_params = {
+            "dim": qwen2_config.get("dim", 896),
+            "layers": 24,  # Default for Qwen2 encoder
+            "heads": 14,
+            "kv_heads": 2,
+            "intermediate_size": 4864,
+            "rms_norm_eps": 1e-6,
+            "rope_theta": 1000000.0,
+        }
+
+        # Update params to include qwen2 config
+        if params.get("params") is None:
+            params["params"] = {}
+        params["params"]["qwen2"] = qwen2_params
+        params["params"]["sam"] = sam_config
+
+        # Set hidden_size from qwen2 dim
+        if "hidden_size" not in params:
+            params["hidden_size"] = qwen2_config.get("dim", 896)
+
+        return super().from_dict(params)
+
+
+@dataclass
+class MLPConfig(BaseModelConfig):
+    hidden_size: int
+    intermediate_size: int
+    hidden_act: str = "gelu"
+
+
+@dataclass
+class ProjectorConfig(BaseModelConfig):
+    projector_type: str = "linear"
+    input_dim: int = 2048
+    n_embed: int = 1280
+    depth: int = 2
+    mlp_ratio: int = 1
+    downsample_ratio: int = 2
+    token_pooling: bool = False
+
+
+@dataclass
+class SAMViTConfig(BaseModelConfig):
+    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 = 128815
+    vision_feature_select_strategy: str = "default"
+    select_layer: int = -1
+    pad_id: int = 100001
+    num_image_tokens: int = 576
+    vocab_size: int = 32000
+    tile_tag: str = "2D"
+    global_view_pos: str = "head"
+    eos_token_id: Optional[List[int]] = None
+    quantization: Optional[Dict] = None
+
+    @classmethod
+    def from_dict(cls, params):
+        if "language_config" in params:
+            params["text_config"] = params["language_config"]
+            del params["language_config"]
+
+        return cls(
+            text_config=TextConfig.from_dict(params["text_config"]),
+            vision_config=VisionConfig.from_dict(params["vision_config"]),
+            projector_config=ProjectorConfig.from_dict(params["projector_config"]),
+            **{
+                k: v
+                for k, v in params.items()
+                if k in inspect.signature(cls).parameters
+                and k not in ["text_config", "vision_config", "projector_config"]
+            },
+        )
+
+
+@dataclass
+class Conversation:
+    """A class that represents a conversation."""
+
+    system: str
+    roles: List[str]
+    messages: List[List[str]]
+    offset: int
+    sep_style: int
+    sep: str
+    sep2: str
+    version: str = "Unknown"
+
+
+@dataclass
+class VLChatProcessorOutput:
+    """
+    Output of the VL chat processor.
+    """
+
+    sft_format: str
+    input_ids: List[int]
+    pixel_values: List
+    num_image_tokens: List[int]
+    image_grid_thw: List[List[int]]
+    image_sizes: Optional[List[List[int]]] = None
+    videos: Optional[List] = None
+    aspect_ratio_ids: Optional[List[int]] = None
+    aspect_ratio_mask: Optional[List[List[int]]] = None
+    cross_attention_mask: Optional[List[List[List[int]]]] = None
+    attention_mask: Optional[List[int]] = None
+    labels: Optional[List[int]] = None
+
+
+@dataclass
+class BatchCollateOutput:
+    input_ids: List
+    labels: List
+    attention_mask: List
+    pixel_values: List