Upload HfMoondream

config.json

@@ -0,0 +1,13 @@
+{
+  "architectures": [
+    "HfMoondream"
+  ],
+  "auto_map": {
+    "AutoConfig": "hf_moondream.HfConfig",
+    "AutoModelForCausalLM": "hf_moondream.HfMoondream"
+  },
+  "config": {},
+  "model_type": "moondream1",
+  "torch_dtype": "float16",
+  "transformers_version": "4.45.2"
+}

config.py

@@ -0,0 +1,86 @@
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional
+
+
+@dataclass(frozen=True)
+class TextConfig:
+    dim: int = 2048
+    ff_dim: int = 8192
+    n_layers: int = 24
+    vocab_size: int = 51200
+    max_context: int = 2048
+    n_heads: int = 32
+    n_kv_heads: int = 32
+    prefix_attn: int = 730
+
+
+@dataclass(frozen=True)
+class VisionConfig:
+    enc_dim: int = 1152
+    enc_patch_size: int = 14
+    enc_n_layers: int = 27
+    enc_ff_dim: int = 4304
+    enc_n_heads: int = 16
+    proj_out_dim: int = 2048
+    crop_size: int = 378
+    in_channels: int = 3
+    max_crops: int = 12
+    overlap_margin: int = 4
+    proj_inner_dim: int = 8192
+
+
+@dataclass(frozen=True)
+class RegionConfig:
+    dim: int = 2048
+    coord_feat_dim: int = 256
+    coord_out_dim: int = 1024
+    size_feat_dim: int = 512
+    size_out_dim: int = 2048
+    inner_dim: int = 8192
+
+
+@dataclass(frozen=True)
+class TokenizerConfig:
+    bos_id: int = 50256
+    eos_id: int = 50256
+    templates: Dict[str, Optional[Dict[str, List[int]]]] = field(
+        default_factory=lambda: {
+            "caption": {
+                "short": [198, 198, 16438, 8305, 25],
+                "normal": [198, 198, 24334, 1159, 25],
+                "long": [198, 198, 14617, 8305, 25],
+            },
+            "query": {"prefix": [198, 198, 24361, 25], "suffix": [198, 198, 33706, 25]},
+            "detect": {"prefix": [198, 198, 47504, 25], "suffix": [628]},
+            "point": {"prefix": [198, 198, 12727, 25], "suffix": [628]},
+        }
+    )
+
+
+@dataclass(frozen=True)
+class MoondreamConfig:
+    text: TextConfig = TextConfig()
+    vision: VisionConfig = VisionConfig()
+    region: RegionConfig = RegionConfig()
+    tokenizer: TokenizerConfig = TokenizerConfig()
+
+    @classmethod
+    def from_dict(cls, config_dict: dict):
+        text_config = TextConfig(**config_dict.get("text", {}))
+        vision_config = VisionConfig(**config_dict.get("vision", {}))
+        region_config = RegionConfig(**config_dict.get("region", {}))
+        tokenizer_config = TokenizerConfig(**config_dict.get("tokenizer", {}))
+        return cls(
+            text=text_config,
+            vision=vision_config,
+            region=region_config,
+            tokenizer=tokenizer_config,
+        )
+
+    def to_dict(self):
+        return {
+            "text": self.text.__dict__,
+            "vision": self.vision.__dict__,
+            "region": self.region.__dict__,
+            "tokenizer": self.tokenizer.__dict__,
+        }

generation_config.json

@@ -0,0 +1,4 @@
+{
+  "_from_model_config": true,
+  "transformers_version": "4.45.2"
+}

hf_moondream.py

@@ -0,0 +1,142 @@
+from transformers import PreTrainedModel, PretrainedConfig
+
+from .config import MoondreamConfig
+from .moondream import MoondreamModel
+
+# Files sometimes don't get loaded without these...
+from .image_crops import *
+from .vision import *
+from .text import *
+from .region import *
+from .utils import *
+
+
+def extract_question(text):
+    prefix = "<image>\n\nQuestion: "
+    suffix = "\n\nAnswer:"
+
+    if text.startswith(prefix) and text.endswith(suffix):
+        return text[len(prefix) : -len(suffix)]
+    else:
+        return None
+
+
+class HfConfig(PretrainedConfig):
+    _auto_class = "AutoConfig"
+    model_type = "moondream1"
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self.config = {}
+
+
+class HfMoondream(PreTrainedModel):
+    _auto_class = "AutoModelForCausalLM"
+    config_class = HfConfig
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.model = MoondreamModel(
+            MoondreamConfig.from_dict(config.config), setup_caches=False
+        )
+        self._is_kv_cache_setup = False
+
+    def _setup_caches(self):
+        if not self._is_kv_cache_setup:
+            self.model._setup_caches()
+            self._is_kv_cache_setup = True
+
+    @property
+    def encode_image(self):
+        self._setup_caches()
+        return self.model.encode_image
+
+    @property
+    def query(self):
+        self._setup_caches()
+        return self.model.query
+
+    @property
+    def caption(self):
+        self._setup_caches()
+        return self.model.caption
+
+    @property
+    def detect(self):
+        self._setup_caches()
+        return self.model.detect
+
+    @property
+    def point(self):
+        self._setup_caches()
+        return self.model.point
+
+    @property
+    def detect_gaze(self):
+        self._setup_caches()
+        return self.model.detect_gaze
+
+    def answer_question(
+        self,
+        image_embeds,
+        question,
+        tokenizer=None,
+        chat_history="",
+        result_queue=None,
+        max_new_tokens=256,
+        **kwargs
+    ):
+        answer = self.query(image_embeds, question)["answer"].strip()
+
+        if result_queue is not None:
+            result_queue.put(answer)
+        return answer
+
+    def batch_answer(self, images, prompts, tokenizer=None, **kwargs):
+        answers = []
+        for image, prompt in zip(images, prompts):
+            answers.append(self.query(image, prompt)["answer"].strip())
+        return answers
+
+    def _unsupported_exception(self):
+        raise NotImplementedError(
+            "This method is not supported in the latest version of moondream. "
+            "Consider upgrading to the updated API spec, or alternately pin "
+            "to 'revision=2024-08-26'."
+        )
+
+    def generate(self, image_embeds, prompt, tokenizer, max_new_tokens=128, **kwargs):
+        """
+        Function definition remains unchanged for backwards compatibility.
+        Be aware that tokenizer, max_new_takens, and kwargs are ignored.
+        """
+        prompt_extracted = extract_question(prompt)
+        if prompt_extracted is not None:
+            answer = self.model.query(
+                image=image_embeds, question=prompt_extracted, stream=False
+            )["answer"]
+        else:
+            image_embeds = self.encode_image(image_embeds)
+            prompt_tokens = torch.tensor(
+                [self.model.tokenizer.encode(prompt).ids],
+                device=self.device,
+            )
+
+            def generator():
+                for token in self.model._generate_text(
+                    prompt_tokens,
+                    image_embeds.kv_cache,
+                    image_embeds.pos,
+                    max_new_tokens,
+                ):
+                    yield token
+
+            answer = "".join(list(generator()))
+
+        return [answer]
+
+    def get_input_embeddings(self):
+        return super().get_input_embeddings()
+
+    def input_embeds(self, *args, **kwargs):
+        self._unsupported_exception()

image_crops.py

@@ -0,0 +1,208 @@
+import math
+import numpy as np
+import torch
+import pyvips
+
+from typing import TypedDict
+
+
+def select_tiling(
+    height: int, width: int, crop_size: int, max_crops: int
+) -> tuple[int, int]:
+    """
+    Determine the optimal number of tiles to cover an image with overlapping crops.
+    """
+    if height <= crop_size or width <= crop_size:
+        return (1, 1)
+
+    # Minimum required tiles in each dimension
+    min_h = math.ceil(height / crop_size)
+    min_w = math.ceil(width / crop_size)
+
+    # If minimum required tiles exceed max_crops, return proportional distribution
+    if min_h * min_w > max_crops:
+        ratio = math.sqrt(max_crops / (min_h * min_w))
+        return (max(1, math.floor(min_h * ratio)), max(1, math.floor(min_w * ratio)))
+
+    # Perfect aspect-ratio tiles that satisfy max_crops
+    h_tiles = math.floor(math.sqrt(max_crops * height / width))
+    w_tiles = math.floor(math.sqrt(max_crops * width / height))
+
+    # Ensure we meet minimum tile requirements
+    h_tiles = max(h_tiles, min_h)
+    w_tiles = max(w_tiles, min_w)
+
+    # If we exceeded max_crops, scale down the larger dimension
+    if h_tiles * w_tiles > max_crops:
+        if w_tiles > h_tiles:
+            w_tiles = math.floor(max_crops / h_tiles)
+        else:
+            h_tiles = math.floor(max_crops / w_tiles)
+
+    return (max(1, h_tiles), max(1, w_tiles))
+
+
+class OverlapCropOutput(TypedDict):
+    crops: np.ndarray
+    tiling: tuple[int, int]
+
+
+def overlap_crop_image(
+    image: np.ndarray,
+    overlap_margin: int,
+    max_crops: int,
+    base_size: tuple[int, int] = (378, 378),
+    patch_size: int = 14,
+) -> OverlapCropOutput:
+    """
+    Process an image using an overlap-and-resize cropping strategy with margin handling.
+
+    This function takes an input image and creates multiple overlapping crops with
+    consistent margins. It produces:
+    1. A single global crop resized to base_size
+    2. Multiple overlapping local crops that maintain high resolution details
+    3. A patch ordering matrix that tracks correspondence between crops
+
+    The overlap strategy ensures:
+    - Smooth transitions between adjacent crops
+    - No loss of information at crop boundaries
+    - Proper handling of features that cross crop boundaries
+    - Consistent patch indexing across the full image
+
+    Args:
+        image (np.ndarray): Input image as numpy array with shape (H,W,C)
+        base_size (tuple[int,int]): Target size for crops, default (378,378)
+        patch_size (int): Size of patches in pixels, default 14
+        overlap_margin (int): Margin size in patch units, default 4
+        max_crops (int): Maximum number of crops allowed, default 12
+
+    Returns:
+        OverlapCropOutput: Dictionary containing:
+            - crops: A numpy array containing the global crop of the full image (index 0)
+                followed by the overlapping cropped regions (indices 1+)
+            - tiling: Tuple of (height,width) tile counts
+    """
+    original_h, original_w = image.shape[:2]
+
+    # Convert margin from patch units to pixels
+    margin_pixels = patch_size * overlap_margin
+    total_margin_pixels = margin_pixels * 2  # Both sides
+
+    # Calculate crop parameters
+    crop_patches = base_size[0] // patch_size  # patches per crop dimension
+    crop_window_patches = crop_patches - (2 * overlap_margin)  # usable patches
+    crop_window_size = crop_window_patches * patch_size  # usable size in pixels
+
+    # Determine tiling
+    tiling = select_tiling(
+        original_h - total_margin_pixels,
+        original_w - total_margin_pixels,
+        crop_window_size,
+        max_crops,
+    )
+
+    # Pre-allocate crops.
+    n_crops = tiling[0] * tiling[1] + 1  # 1 = global crop
+    crops = np.zeros(
+        (n_crops, base_size[0], base_size[1], image.shape[2]), dtype=np.uint8
+    )
+
+    # Resize image to fit tiling
+    target_size = (
+        tiling[0] * crop_window_size + total_margin_pixels,
+        tiling[1] * crop_window_size + total_margin_pixels,
+    )
+
+    # Convert to vips for resizing
+    vips_image = pyvips.Image.new_from_array(image)
+    scale_x = target_size[1] / image.shape[1]
+    scale_y = target_size[0] / image.shape[0]
+    resized = vips_image.resize(scale_x, vscale=scale_y)
+    image = resized.numpy()
+
+    # Create global crop
+    scale_x = base_size[1] / vips_image.width
+    scale_y = base_size[0] / vips_image.height
+    global_vips = vips_image.resize(scale_x, vscale=scale_y)
+    crops[0] = global_vips.numpy()
+
+    for i in range(tiling[0]):
+        for j in range(tiling[1]):
+            # Calculate crop coordinates
+            y0 = i * crop_window_size
+            x0 = j * crop_window_size
+
+            # Extract crop with padding if needed
+            y_end = min(y0 + base_size[0], image.shape[0])
+            x_end = min(x0 + base_size[1], image.shape[1])
+
+            crop_region = image[y0:y_end, x0:x_end]
+            crops[
+                1 + i * tiling[1] + j, : crop_region.shape[0], : crop_region.shape[1]
+            ] = crop_region
+
+    return {"crops": crops, "tiling": tiling}
+
+
+def reconstruct_from_crops(
+    crops: torch.Tensor,
+    tiling: tuple[int, int],
+    overlap_margin: int,
+    patch_size: int = 14,
+) -> torch.Tensor:
+    """
+    Reconstruct the original image from overlapping crops into a single seamless image.
+
+    Takes a list of overlapping image crops along with their positional metadata and
+    reconstructs them into a single coherent image by carefully stitching together
+    non-overlapping regions. Handles both numpy arrays and PyTorch tensors.
+
+    Args:
+        crops: List of image crops as numpy arrays or PyTorch tensors with shape
+            (H,W,C)
+        tiling: Tuple of (height,width) indicating crop grid layout
+        patch_size: Size in pixels of each patch, default 14
+        overlap_margin: Number of overlapping patches on each edge, default 4
+
+    Returns:
+        Reconstructed image as numpy array or PyTorch tensor matching input type,
+        with shape (H,W,C) where H,W are the original image dimensions
+    """
+    tiling_h, tiling_w = tiling
+    crop_height, crop_width = crops[0].shape[:2]
+    margin_pixels = overlap_margin * patch_size
+
+    # Calculate output size (only adding margins once)
+    output_h = (crop_height - 2 * margin_pixels) * tiling_h + 2 * margin_pixels
+    output_w = (crop_width - 2 * margin_pixels) * tiling_w + 2 * margin_pixels
+
+    reconstructed = torch.zeros(
+        (output_h, output_w, crops[0].shape[2]),
+        device=crops[0].device,
+        dtype=crops[0].dtype,
+    )
+
+    for i, crop in enumerate(crops):
+        tile_y = i // tiling_w
+        tile_x = i % tiling_w
+
+        # For each tile, determine which part to keep
+        # Keep left margin only for first column
+        x_start = 0 if tile_x == 0 else margin_pixels
+        # Keep right margin only for last column
+        x_end = crop_width if tile_x == tiling_w - 1 else crop_width - margin_pixels
+        # Keep top margin only for first row
+        y_start = 0 if tile_y == 0 else margin_pixels
+        # Keep bottom margin only for last row
+        y_end = crop_height if tile_y == tiling_h - 1 else crop_height - margin_pixels
+
+        # Calculate where this piece belongs in the output
+        out_x = tile_x * (crop_width - 2 * margin_pixels)
+        out_y = tile_y * (crop_height - 2 * margin_pixels)
+
+        # Place the piece
+        reconstructed[
+            out_y + y_start : out_y + y_end, out_x + x_start : out_x + x_end
+        ] = crop[y_start:y_end, x_start:x_end]
+
+    return reconstructed

layers.py

@@ -0,0 +1,63 @@
+from dataclasses import dataclass
+from typing import Literal
+
+import torch
+from torch.nn import functional as F
+
+
+def gelu_approx(x):
+    return F.gelu(x, approximate="tanh")
+
+
+@dataclass
+class LinearWeights:
+    weight: torch.Tensor
+    bias: torch.Tensor
+
+
+def linear(x: torch.Tensor, w: LinearWeights) -> torch.Tensor:
+    return F.linear(x, w.weight, w.bias)
+
+
+@dataclass
+class LayerNormWeights:
+    weight: torch.Tensor
+    bias: torch.Tensor
+
+
+def layer_norm(x: torch.Tensor, w: LayerNormWeights) -> torch.Tensor:
+    return F.layer_norm(x, w.bias.shape, w.weight, w.bias)
+
+
+@dataclass
+class MLPWeights:
+    fc1: LinearWeights
+    fc2: LinearWeights
+    act: Literal["gelu_approx"] = "gelu_approx"
+
+
+def mlp(x: torch.Tensor, w: MLPWeights) -> torch.Tensor:
+    x = w.fc1(x)
+    x = gelu_approx(x)
+    x = w.fc2(x)
+    return x
+
+
+@dataclass
+class AttentionWeights:
+    qkv: LinearWeights
+    proj: LinearWeights
+
+
+def attn(x: torch.Tensor, w: AttentionWeights, n_heads: int) -> torch.Tensor:
+    bsz, q_len, d_model = x.shape
+    head_dim = d_model // n_heads
+
+    q, k, v = [
+        t.view(bsz, q_len, n_heads, head_dim).transpose(1, 2)
+        for t in linear(x, w.qkv).chunk(3, dim=-1)
+    ]
+    out = F.scaled_dot_product_attention(q, k, v)
+    out = out.transpose(1, 2).reshape(bsz, q_len, d_model)
+    out = linear(out, w.proj)
+    return out

model.safetensors

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:23e2e6498a058d12832e119dc97a1d2f14936b4ccf77b8492bc0fefba49ea8bb
+size 3854538376

moondream.py

@@ -0,0 +1,717 @@
+import torch
+import torch.nn as nn
+import random
+
+from typing import Literal, Tuple, TypedDict, Union, Dict, Any, Optional, List
+from PIL import Image
+from dataclasses import dataclass
+from tokenizers import Tokenizer
+
+from .config import MoondreamConfig
+from .image_crops import reconstruct_from_crops
+from .vision import vision_encoder, vision_projection, prepare_crops, build_vision_model
+from .text import build_text_model, text_encoder, lm_head, text_decoder
+from .region import decode_coordinate, encode_coordinate, decode_size, encode_size
+from .utils import remove_outlier_points
+
+
+TextSamplingSettings = TypedDict(
+    "TextSamplingSettings",
+    {
+        "max_tokens": int,
+        "temperature": float,
+        "top_p": float,
+    },
+    total=False,
+)
+
+ObjectSamplingSettings = TypedDict(
+    "ObjectSamplingSettings",
+    {"max_objects": int},
+    total=False,
+)
+
+DEFAULT_MAX_TOKENS = 768
+DEFAULT_TEMPERATURE = 0.5
+DEFAULT_TOP_P = 0.3
+DEFAULT_MAX_OBJECTS = 50
+
+
+@dataclass(frozen=True)
+class EncodedImage:
+    pos: int
+    caches: List[Tuple[torch.Tensor, torch.Tensor]]
+
+
+class KVCache(nn.Module):
+
+    def __init__(self, n_heads, n_kv_heads, max_context, dim, device, dtype):
+        super().__init__()
+        cache_shape = (1, n_kv_heads, max_context, dim // n_heads)
+        self.register_buffer(
+            "k_cache", torch.zeros(*cache_shape, device=device, dtype=dtype)
+        )
+        self.register_buffer(
+            "v_cache", torch.zeros(*cache_shape, device=device, dtype=dtype)
+        )
+
+    def update(self, pos_ids, k, v):
+        kout, vout = self.k_cache, self.v_cache
+        kout[:, :, pos_ids, :] = k
+        vout[:, :, pos_ids, :] = v
+        return kout, vout
+
+
+class MoondreamModel(nn.Module):
+    def __init__(self, config: MoondreamConfig, dtype=torch.float16, setup_caches=True):
+        super().__init__()
+        self.config = config
+
+        self.tokenizer = Tokenizer.from_pretrained(
+            "vikhyatk/moondream2", revision="2025-01-09"
+        )
+        self.vision = build_vision_model(config.vision, dtype)
+        self.text = build_text_model(config.text, dtype)
+
+        # Region Model
+        self.region = nn.ModuleDict(
+            {
+                "coord_encoder": nn.Linear(
+                    config.region.coord_feat_dim, config.region.dim, dtype=dtype
+                ),
+                "coord_decoder": nn.ModuleDict(
+                    {
+                        "fc1": nn.Linear(
+                            config.region.dim, config.region.inner_dim, dtype=dtype
+                        ),
+                        "fc2": nn.Linear(
+                            config.region.inner_dim,
+                            config.region.coord_out_dim,
+                            dtype=dtype,
+                        ),
+                    }
+                ),
+                "size_encoder": nn.Linear(
+                    config.region.size_feat_dim, config.region.dim, dtype=dtype
+                ),
+                "size_decoder": nn.ModuleDict(
+                    {
+                        "fc1": nn.Linear(
+                            config.region.dim, config.region.inner_dim, dtype=dtype
+                        ),
+                        "fc2": nn.Linear(
+                            config.region.inner_dim,
+                            config.region.size_out_dim,
+                            dtype=dtype,
+                        ),
+                    }
+                ),
+            }
+        )
+        self.region.coord_features = nn.Parameter(
+            torch.empty(config.region.coord_feat_dim // 2, 1, dtype=dtype).T
+        )
+        self.region.size_features = nn.Parameter(
+            torch.empty(config.region.size_feat_dim // 2, 2, dtype=dtype).T
+        )
+
+        attn_mask = torch.tril(
+            torch.ones(
+                1, 1, config.text.max_context, config.text.max_context, dtype=torch.bool
+            )
+        )
+        patch_w = config.vision.crop_size // config.vision.enc_patch_size
+        prefix_attn_len = 1 + patch_w**2
+        attn_mask[..., :prefix_attn_len, :prefix_attn_len] = 1
+        self.register_buffer("attn_mask", attn_mask, persistent=False)
+
+        # Initialize KV caches.
+        if setup_caches:
+            self._setup_caches()
+
+    def _setup_caches(self):
+        c = self.config.text
+        for b in self.text.blocks:
+            b.kv_cache = KVCache(
+                c.n_heads,
+                c.n_kv_heads,
+                c.max_context,
+                c.dim,
+                device=self.device,
+                dtype=self.vision.pos_emb.dtype,
+            )
+
+    @property
+    def device(self):
+        return self.vision.pos_emb.device
+
+    def _vis_enc(self, x: torch.Tensor):
+        return vision_encoder(x, self.vision, self.config.vision)
+
+    def _vis_proj(self, g: torch.Tensor, r: torch.Tensor):
+        return vision_projection(g, r, self.vision, self.config.vision)
+
+    def _prefill(self, x: torch.Tensor, attn_mask: torch.Tensor, pos_ids: torch.Tensor):
+        return text_decoder(x, self.text, attn_mask, pos_ids, self.config.text)
+
+    def _decode_one_tok(
+        self, x: torch.Tensor, attn_mask: torch.Tensor, pos_ids: torch.Tensor
+    ):
+        hidden = text_decoder(x, self.text, attn_mask, pos_ids, self.config.text)
+        logits = lm_head(hidden, self.text)
+        return logits, hidden
+
+    def compile(self):
+        # TODO: vision_projection is not being compiled
+        self._vis_enc = torch.compile(self._vis_enc, fullgraph=True)
+        self._prefill = torch.compile(self._prefill, fullgraph=True)
+        self._decode_one_tok = torch.compile(
+            self._decode_one_tok, fullgraph=True, mode="reduce-overhead"
+        )
+
+    def _run_vision_encoder(self, image: Image.Image) -> torch.Tensor:
+        all_crops, tiling = prepare_crops(image, self.config.vision, device=self.device)
+        torch._dynamo.mark_dynamic(all_crops, 0)
+
+        outputs = self._vis_enc(all_crops)
+
+        global_features = outputs[0]
+        local_features = outputs[1:].view(
+            -1,
+            self.config.vision.enc_n_layers,
+            self.config.vision.enc_n_layers,
+            self.config.vision.enc_dim,
+        )
+
+        reconstructed = reconstruct_from_crops(
+            local_features,
+            tiling,
+            patch_size=1,
+            overlap_margin=self.config.vision.overlap_margin,
+        )
+
+        return self._vis_proj(global_features, reconstructed)
+
+    def encode_image(self, image: Union[Image.Image, EncodedImage]) -> EncodedImage:
+        if isinstance(image, EncodedImage):
+            return image
+        elif not isinstance(image, Image.Image):
+            raise ValueError("image must be a PIL Image or EncodedImage")
+
+        # Run through text model in addition to the vision encoder, to minimize
+        # re-computation if multiple queries are performed on this image.
+        with torch.inference_mode():
+            img_emb = self._run_vision_encoder(image)
+            bos_emb = text_encoder(
+                torch.tensor([[self.config.tokenizer.bos_id]], device=self.device),
+                self.text,
+            )
+            inputs_embeds = torch.cat([bos_emb, img_emb[None]], dim=1)
+            mask = self.attn_mask[:, :, 0 : inputs_embeds.size(1), :]
+            pos_ids = torch.arange(inputs_embeds.size(1), dtype=torch.long)
+            self._prefill(inputs_embeds, mask, pos_ids)
+
+        return EncodedImage(
+            pos=inputs_embeds.size(1),
+            caches=[
+                (
+                    b.kv_cache.k_cache[:, :, : inputs_embeds.size(1), :].clone(),
+                    b.kv_cache.v_cache[:, :, : inputs_embeds.size(1), :].clone(),
+                )
+                for b in self.text.blocks
+            ],
+        )
+
+    def _apply_top_p(self, probs: torch.Tensor, top_p: float):
+        probs_sort, probs_idx = torch.sort(probs, dim=-1, descending=True)
+        probs_sum = torch.cumsum(probs_sort, dim=-1)
+        mask = probs_sum - probs_sort > top_p
+        probs_sort[mask] = 0.0
+        probs_sort.div_(probs_sort.sum(dim=-1, keepdim=True))
+        next_probs = torch.zeros_like(probs)
+        next_probs.scatter_(dim=-1, index=probs_idx, src=probs_sort)
+        return next_probs
+
+    def _prefill_prompt(
+        self, prompt_tokens: torch.Tensor, pos: int, temperature: float, top_p: float
+    ):
+        with torch.inference_mode():
+            prompt_emb = text_encoder(prompt_tokens, self.text)
+            torch._dynamo.mark_dynamic(prompt_emb, 1)
+            mask = self.attn_mask[:, :, pos : pos + prompt_emb.size(1), :]
+            pos_ids = torch.arange(pos, pos + prompt_emb.size(1), dtype=torch.long)
+            hidden = self._prefill(prompt_emb, mask, pos_ids)
+            logits = lm_head(hidden, self.text)
+
+            if temperature == 0:
+                next_token = torch.argmax(logits, dim=-1).unsqueeze(1)
+            else:
+                probs = torch.softmax(logits / temperature, dim=-1)
+                probs = self._apply_top_p(probs, top_p)
+                next_token = torch.multinomial(probs, num_samples=1)
+
+        pos = pos + prompt_emb.size(1)
+        return logits, hidden, next_token, pos
+
+    def _generate_text(
+        self,
+        prompt_tokens: torch.Tensor,
+        pos: int,
+        settings: Optional[TextSamplingSettings] = None,
+    ):
+        max_tokens = (
+            settings.get("max_tokens", DEFAULT_MAX_TOKENS)
+            if settings
+            else DEFAULT_MAX_TOKENS
+        )
+        temperature = (
+            settings.get("temperature", DEFAULT_TEMPERATURE)
+            if settings
+            else DEFAULT_TEMPERATURE
+        )
+        top_p = settings.get("top_p", DEFAULT_TOP_P) if settings else DEFAULT_TOP_P
+
+        _, _, next_token, pos = self._prefill_prompt(
+            prompt_tokens, pos, temperature, top_p
+        )
+
+        def generator(next_token, pos):
+            mask = torch.zeros(1, 1, 2048, device=self.device, dtype=torch.bool)
+            mask[:, :, :pos] = 1
+            pos_ids = torch.tensor([pos], device=self.device, dtype=torch.long)
+            generated_tokens = 0
+
+            # For properly handling token streaming with Unicode
+            token_cache = []
+            print_len = 0
+
+            while (
+                next_token_id := next_token.item()
+            ) != self.config.tokenizer.eos_id and generated_tokens < max_tokens:
+                # Add token to our cache
+                token_cache.append(next_token_id)
+
+                # Decode all tokens collected so far
+                text = self.tokenizer.decode(token_cache)
+
+                # After a newline, we flush the cache completely
+                if text.endswith("\n"):
+                    printable_text = text[print_len:]
+                    token_cache = []
+                    print_len = 0
+                    if printable_text:
+                        yield printable_text
+                # If the last token is a CJK character, we can safely print it
+                elif len(text) > 0 and _is_cjk_char(ord(text[-1])):
+                    printable_text = text[print_len:]
+                    print_len += len(printable_text)
+                    if printable_text:
+                        yield printable_text
+                # Otherwise, only print up to the last space to avoid cutting words
+                else:
+                    last_space_idx = text.rfind(" ", print_len)
+                    if last_space_idx >= print_len:
+                        printable_text = text[print_len : last_space_idx + 1]
+                        print_len += len(printable_text)
+                        if printable_text:
+                            yield printable_text
+
+                with torch.inference_mode():
+                    next_emb = text_encoder(next_token, self.text)
+                    mask[:, :, pos], pos_ids[0] = 1, pos
+                    logits, _ = self._decode_one_tok(next_emb, mask, pos_ids)
+                    pos += 1
+
+                    if temperature == 0:
+                        next_token = torch.argmax(logits, dim=-1).unsqueeze(1)  # (1, 1)
+                    else:
+                        probs = torch.softmax(logits / temperature, dim=-1)  # (1, V)
+                        probs = self._apply_top_p(probs, top_p)
+                        next_token = torch.multinomial(probs, num_samples=1)  # (1, 1)
+
+                    generated_tokens += 1
+
+            # Flush any remaining text in the cache
+            if token_cache:
+                text = self.tokenizer.decode(token_cache)
+                printable_text = text[print_len:]
+                if printable_text:
+                    yield printable_text
+
+        return generator(next_token, pos)
+
+    def query(
+        self,
+        image: Union[Image.Image, EncodedImage],
+        question: str,
+        stream: bool = False,
+        settings: Optional[TextSamplingSettings] = None,
+    ):
+        if self.config.tokenizer.templates["query"] is None:
+            raise NotImplementedError("Model does not support querying.")
+
+        image = self.encode_image(image)
+        self.load_encoded_image(image)
+
+        prompt_tokens = torch.tensor(
+            [
+                self.config.tokenizer.templates["query"]["prefix"]
+                + self.tokenizer.encode(" " + question).ids
+                + self.config.tokenizer.templates["query"]["suffix"]
+            ],
+            device=self.device,
+        )
+
+        def generator():
+            for token in self._generate_text(prompt_tokens, image.pos, settings):
+                yield token
+
+        if stream:
+            return {"answer": generator()}
+        else:
+            return {"answer": "".join(list(generator()))}
+
+    def load_encoded_image(self, encoded_image: EncodedImage):
+        for b, (k, v) in zip(self.text.blocks, encoded_image.caches):
+            b.kv_cache.k_cache[:, :, : k.size(2), :] = k
+            b.kv_cache.v_cache[:, :, : v.size(2), :] = v
+
+    def caption(
+        self,
+        image: Union[Image.Image, EncodedImage],
+        length: Literal["normal", "short", "long"] = "normal",
+        stream: bool = False,
+        settings: Optional[TextSamplingSettings] = None,
+    ):
+        if self.config.tokenizer.templates["caption"] is None:
+            raise NotImplementedError("Model does not support captioning.")
+        if length not in self.config.tokenizer.templates["caption"]:
+            raise ValueError(f"Model does not support caption length '{length}'.")
+
+        image = self.encode_image(image)
+        self.load_encoded_image(image)
+
+        prompt_tokens = torch.tensor(
+            [self.config.tokenizer.templates["caption"][length]], device=self.device
+        )
+
+        def generator():
+            for token in self._generate_text(prompt_tokens, image.pos, settings):
+                yield token
+
+        if stream:
+            return {"caption": generator()}
+        else:
+            return {"caption": "".join(list(generator()))}
+
+    def _generate_points(
+        self,
+        hidden: torch.Tensor,
+        next_token: torch.Tensor,
+        pos: int,
+        include_size: bool = True,
+        max_objects: int = DEFAULT_MAX_OBJECTS,
+    ):
+        out = []
+        mask = torch.zeros(1, 1, 2048, device=self.device, dtype=torch.bool)
+        mask[:, :, :pos] = 1
+        pos_ids = torch.tensor([pos], device=self.device, dtype=torch.long)
+
+        with torch.inference_mode():
+            while (
+                next_token.item() != self.config.tokenizer.eos_id
+                and len(out) < max_objects
+            ):
+                x_logits = decode_coordinate(hidden, self.region)
+                x_center = torch.argmax(x_logits, dim=-1) / x_logits.size(-1)
+                next_emb = encode_coordinate(
+                    x_center.to(dtype=x_logits.dtype), self.region
+                ).unsqueeze(0)
+
+                # Decode y-coordinate
+                mask[:, :, pos], pos_ids[0] = 1, pos
+                _, hidden = self._decode_one_tok(next_emb, mask, pos_ids)
+                pos += 1
+                y_logits = decode_coordinate(hidden, self.region)
+                y_center = torch.argmax(y_logits, dim=-1) / y_logits.size(-1)
+                next_emb = encode_coordinate(
+                    y_center.to(dtype=y_logits.dtype), self.region
+                ).unsqueeze(0)
+
+                # Decode size
+                if include_size:
+                    mask[:, :, pos], pos_ids[0] = 1, pos
+                    logits, hidden = self._decode_one_tok(next_emb, mask, pos_ids)
+                    pos += 1
+                    size_logits = decode_size(hidden, self.region)
+
+                    # Get bin indices from the logits
+                    w_bin = torch.argmax(size_logits[0], dim=-1)
+                    h_bin = torch.argmax(size_logits[1], dim=-1)
+
+                    # Convert from bin indices to actual size values using the inverse of the log-scale mapping
+                    # Formula: size = 2^((bin / 1023.0) * 10.0 - 10.0)
+                    w = torch.pow(2.0, (w_bin.float() / 1023.0) * 10.0 - 10.0)
+                    h = torch.pow(2.0, (h_bin.float() / 1023.0) * 10.0 - 10.0)
+
+                    next_emb = (
+                        encode_size(
+                            torch.tensor(
+                                [w, h], device=self.device, dtype=size_logits.dtype
+                            ),
+                            self.region,
+                        )
+                        .unsqueeze(0)
+                        .unsqueeze(0)
+                    )
+
+                    # Add object
+                    out.append(
+                        {
+                            "x_min": x_center.item() - w.item() / 2,
+                            "y_min": y_center.item() - h.item() / 2,
+                            "x_max": x_center.item() + w.item() / 2,
+                            "y_max": y_center.item() + h.item() / 2,
+                        }
+                    )
+                else:
+                    out.append({"x": x_center.item(), "y": y_center.item()})
+
+                # Decode next token (x-coordinate, or eos)
+                mask[:, :, pos], pos_ids[0] = 1, pos
+                logits, hidden = self._decode_one_tok(next_emb, mask, pos_ids)
+                pos += 1
+                next_token = torch.argmax(logits, dim=-1)
+
+        return out
+
+    def detect(
+        self,
+        image: Union[Image.Image, EncodedImage],
+        object: str,
+        settings: Optional[ObjectSamplingSettings] = None,
+    ):
+        if self.config.tokenizer.templates["detect"] is None:
+            raise NotImplementedError("Model does not support object detection.")
+
+        image = self.encode_image(image)
+        self.load_encoded_image(image)
+
+        prompt_tokens = torch.tensor(
+            [
+                self.config.tokenizer.templates["detect"]["prefix"]
+                + self.tokenizer.encode(" " + object).ids
+                + self.config.tokenizer.templates["detect"]["suffix"]
+            ],
+            device=self.device,
+        )
+
+        _, hidden, next_token, pos = self._prefill_prompt(
+            prompt_tokens, image.pos, temperature=0, top_p=0
+        )
+        hidden = hidden[:, -1:, :]
+
+        max_objects = (
+            settings.get("max_objects", DEFAULT_MAX_OBJECTS)
+            if settings
+            else DEFAULT_MAX_OBJECTS
+        )
+        objects = self._generate_points(
+            hidden, next_token, pos, include_size=True, max_objects=max_objects
+        )
+
+        return {"objects": objects}
+
+    def point(
+        self,
+        image: Union[Image.Image, EncodedImage],
+        object: str,
+        settings: Optional[ObjectSamplingSettings] = None,
+    ):
+        if self.config.tokenizer.templates["point"] is None:
+            raise NotImplementedError("Model does not support pointing.")
+
+        image = self.encode_image(image)
+        self.load_encoded_image(image)
+
+        prompt_tokens = torch.tensor(
+            [
+                self.config.tokenizer.templates["point"]["prefix"]
+                + self.tokenizer.encode(" " + object).ids
+                + self.config.tokenizer.templates["point"]["suffix"]
+            ],
+            device=self.device,
+        )
+
+        _, hidden, next_token, pos = self._prefill_prompt(
+            prompt_tokens, image.pos, temperature=0, top_p=0
+        )
+        hidden = hidden[:, -1:, :]
+
+        max_objects = (
+            settings.get("max_objects", DEFAULT_MAX_OBJECTS)
+            if settings
+            else DEFAULT_MAX_OBJECTS
+        )
+        objects = self._generate_points(
+            hidden, next_token, pos, include_size=False, max_objects=max_objects
+        )
+
+        return {"points": objects}
+
+    def _detect_gaze(
+        self,
+        image: EncodedImage,
+        source: Tuple[float, float],
+        force_detect: bool = False,
+    ):
+        with torch.inference_mode():
+            before_emb = text_encoder(
+                torch.tensor(
+                    [self.tokenizer.encode("\n\nPoint:").ids], device=self.device
+                ),
+                self.text,
+            )
+            after_emb = text_encoder(
+                torch.tensor(
+                    [self.tokenizer.encode(" gaze\n\n").ids], device=self.device
+                ),
+                self.text,
+            )
+            x_emb = encode_coordinate(
+                torch.tensor([[[source[0]]]], device=self.device, dtype=torch.float16),
+                self.region,
+            )
+            y_emb = encode_coordinate(
+                torch.tensor([[[source[1]]]], device=self.device, dtype=torch.float16),
+                self.region,
+            )
+
+            prompt_emb = torch.cat([before_emb, x_emb, y_emb, after_emb], dim=1)
+
+            self.load_encoded_image(image)
+
+            mask = self.attn_mask[:, :, image.pos : image.pos + prompt_emb.size(1), :]
+            pos_ids = torch.arange(
+                image.pos, image.pos + prompt_emb.size(1), dtype=torch.long
+            )
+            hidden = self._prefill(prompt_emb, mask, pos_ids)
+            logits = lm_head(hidden, self.text)
+            next_token = torch.argmax(logits, dim=-1)
+            pos = image.pos + prompt_emb.size(1)
+            hidden = hidden[:, -1:, :]
+
+            if force_detect:
+                next_token = torch.tensor([[0]], device=self.device)
+
+            if next_token.item() == self.config.tokenizer.eos_id:
+                return None
+
+            gaze = self._generate_points(
+                hidden, next_token, pos, include_size=False, max_objects=1
+            )
+            return gaze[0]
+
+    def detect_gaze(
+        self,
+        image: Union[Image.Image, EncodedImage],
+        eye: Optional[Tuple[float, float]] = None,
+        face: Optional[Dict[str, float]] = None,
+        unstable_settings: Dict[str, Any] = {},
+    ):
+        if "force_detect" in unstable_settings:
+            force_detect = unstable_settings["force_detect"]
+        else:
+            force_detect = False
+
+        if "prioritize_accuracy" in unstable_settings:
+            prioritize_accuracy = unstable_settings["prioritize_accuracy"]
+        else:
+            prioritize_accuracy = False
+
+        if not prioritize_accuracy:
+            if eye is None:
+                raise ValueError("eye must be provided when prioritize_accuracy=False")
+            image = self.encode_image(image)
+            return {"gaze": self._detect_gaze(image, eye, force_detect=force_detect)}
+        else:
+            if (
+                not isinstance(image, Image.Image)
+                and "flip_enc_img" not in unstable_settings
+            ):
+                raise ValueError(
+                    "image must be a PIL Image when prioritize_accuracy=True, "
+                    "or flip_enc_img must be provided"
+                )
+            if face is None:
+                raise ValueError("face must be provided when prioritize_accuracy=True")
+
+            encoded_image = self.encode_image(image)
+            if (
+                isinstance(image, Image.Image)
+                and "flip_enc_img" not in unstable_settings
+            ):
+                flipped_pil = image.copy()
+                flipped_pil = flipped_pil.transpose(method=Image.FLIP_LEFT_RIGHT)
+                encoded_flipped_image = self.encode_image(flipped_pil)
+            else:
+                encoded_flipped_image = unstable_settings["flip_enc_img"]
+
+            N = 10
+
+            detections = [
+                self._detect_gaze(
+                    encoded_image,
+                    (
+                        random.uniform(face["x_min"], face["x_max"]),
+                        random.uniform(face["y_min"], face["y_max"]),
+                    ),
+                    force_detect=force_detect,
+                )
+                for _ in range(N)
+            ]
+            detections = [
+                (gaze["x"], gaze["y"]) for gaze in detections if gaze is not None
+            ]
+            flipped_detections = [
+                self._detect_gaze(
+                    encoded_flipped_image,
+                    (
+                        1 - random.uniform(face["x_min"], face["x_max"]),
+                        random.uniform(face["y_min"], face["y_max"]),
+                    ),
+                    force_detect=force_detect,
+                )
+                for _ in range(N)
+            ]
+            detections.extend(
+                [
+                    (1 - gaze["x"], gaze["y"])
+                    for gaze in flipped_detections
+                    if gaze is not None
+                ]
+            )
+
+            if len(detections) < N:
+                return {"gaze": None}
+
+            detections = remove_outlier_points(detections)
+            mean_gaze = (
+                sum(gaze[0] for gaze in detections) / len(detections),
+                sum(gaze[1] for gaze in detections) / len(detections),
+            )
+
+            return {"gaze": {"x": mean_gaze[0], "y": mean_gaze[1]}}
+
+
+def _is_cjk_char(cp):
+    """Checks whether CP is the codepoint of a CJK character."""
+    # This defines a "chinese character" as anything in the CJK Unicode block:
+    # https://en.wikipedia.org/wiki/CJK_Unified_Ideographs_(Unicode_block)
+    if (
+        (cp >= 0x4E00 and cp <= 0x9FFF)
+        or (cp >= 0x3400 and cp <= 0x4DBF)
+        or (cp >= 0x2F800 and cp <= 0x2FA1F)
+    ):
+        return True
+    return False

region.py

@@ -0,0 +1,89 @@
+import torch
+import torch.nn as nn
+import math
+
+from .layers import linear, mlp
+
+
+def fourier_features(x: torch.Tensor, w: torch.Tensor) -> torch.Tensor:
+    """
+    Applies Fourier feature mapping to input tensor x using frequency matrix w. This
+    projects inputs through sinusoidal functions to create higher dimensional features
+    that help mitigate spectral bias - the tendency of neural networks to learn
+    low-frequency functions more easily than high-frequency ones. By explicitly
+    mapping inputs to higher frequencies through sin/cos transformations, we enable
+    better learning of fine details and higher frequency patterns.
+
+    Args:
+        x: Input tensor to transform
+        w: Matrix of frequencies for the Fourier features transformation
+
+    Returns:
+        Concatenated cosine and sine transformed features as a tensor
+    """
+    f = 2 * math.pi * x @ w
+    return torch.cat([f.cos(), f.sin()], dim=-1)
+
+
+def encode_coordinate(coord: torch.Tensor, w: nn.Module) -> torch.Tensor:
+    """
+    Takes as input a tensor containing a single float coordinate value (x or y)
+    and encodes it into hidden states for input to the text model.
+
+    Args:
+        coord: Tensor with single float coordinate value
+
+    Returns:
+        Encoded hidden states tensor for input to text model
+    """
+    return linear(fourier_features(coord, w.coord_features), w.coord_encoder)
+
+
+def decode_coordinate(hidden_state: torch.Tensor, w: nn.Module) -> torch.Tensor:
+    """
+    Takes as input the last hidden state from the text model and outputs a single logit
+    representing either an x or y coordinate prediction.
+
+    Args:
+        hidden_state: The final hidden state tensor from the text model.
+
+    Returns:
+        A single logit representing the predicted coordinate value (x or y)
+    """
+    return mlp(hidden_state, w.coord_decoder)
+
+
+def encode_size(size: torch.Tensor, w: nn.Module) -> torch.Tensor:
+    """
+    Takes a tensor containing width and height values and encodes them into
+    hidden states for input to the text model.
+
+    Args:
+        size: Tensor with two floats for width and height
+
+    Returns:
+        Encoded hidden states tensor for input to text model
+    """
+    return linear(fourier_features(size, w.size_features), w.size_encoder)
+
+
+def decode_size(hidden_state: torch.Tensor, w: nn.Module) -> torch.Tensor:
+    """
+    Takes as input the last hidden state from the text model and outputs logits
+    for 1024 bins representing width and height in log-scale.
+
+    The bins are distributed according to the formula:
+    bin = (log2(size) + 10.0) / 10.0 * 1023.0
+    where size values are clamped to be at least 1/1024.
+
+    To convert from bin back to size:
+    size = 2^((bin / 1023.0) * 10.0 - 10.0)
+
+    Args:
+        hidden_state: The final hidden state tensor from the text model.
+
+    Returns:
+        A tensor containing logits for 1024 bins for width and height.
+        Shape is (2, 1024) where the first dimension corresponds to width and height.
+    """
+    return mlp(hidden_state, w.size_decoder).view(2, -1)

rope.py

@@ -0,0 +1,48 @@
+# Ethically sourced from https://github.com/xjdr-alt/entropix
+
+import torch
+
+
+def precompute_freqs_cis(
+    dim: int,
+    end: int,
+    theta: float = 10000.0,
+    use_scaled: bool = False,
+    dtype: torch.dtype = torch.float32,
+) -> torch.Tensor:
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=dtype)[: (dim // 2)] / dim))
+    t = torch.arange(end, dtype=dtype).unsqueeze(1)
+    freqs = t * freqs.unsqueeze(0)
+    freqs = torch.exp(1j * freqs)
+    return torch.stack([freqs.real, freqs.imag], dim=-1)
+
+
+def apply_rotary_emb(
+    x: torch.Tensor,
+    freqs_cis: torch.Tensor,
+    position_ids: torch.Tensor,
+    num_heads: int,
+    rot_dim: int = 32,
+    interleave: bool = False,
+) -> torch.Tensor:
+    assert rot_dim == freqs_cis.shape[-2] * 2
+    assert num_heads == x.shape[1]
+
+    x_rot, x_pass = x[..., :rot_dim], x[..., rot_dim:]
+
+    if interleave:
+        xq_r = x_rot.float().reshape(*x_rot.shape[:-1], -1, 2)[..., 0]
+        xq_i = x_rot.float().reshape(*x_rot.shape[:-1], -1, 2)[..., 1]
+    else:
+        d_q = x_rot.shape[-1] // 2
+        xq_r, xq_i = x_rot[..., :d_q], x_rot[..., d_q:]
+
+    freqs_cos = freqs_cis[..., 0][position_ids, :].unsqueeze(0).unsqueeze(0)
+    freqs_sin = freqs_cis[..., 1][position_ids, :].unsqueeze(0).unsqueeze(0)
+
+    # Complex multiplication: (a + bi) * (c + di) = (ac - bd) + (ad + bc)i
+    xq_out_r = xq_r * freqs_cos - xq_i * freqs_sin
+    xq_out_i = xq_r * freqs_sin + xq_i * freqs_cos
+    xq_out = torch.stack((xq_out_r, xq_out_i), dim=-1).flatten(-2)
+
+    return torch.cat([xq_out.to(x.dtype), x_pass], dim=-1)

text.py

@@ -0,0 +1,205 @@
+import torch
+import torch.nn as nn
+
+from torch.nn import functional as F
+
+from .layers import layer_norm, mlp
+from .rope import apply_rotary_emb, precompute_freqs_cis
+from .config import TextConfig
+
+
+def text_encoder(input_ids: torch.Tensor, w: nn.Module):
+    return F.embedding(input_ids, w.wte)
+
+
+def attn(
+    x: torch.Tensor,
+    w: nn.Module,
+    freqs_cis: torch.Tensor,
+    kv_cache: nn.Module,
+    attn_mask: torch.Tensor,
+    n_heads: int,
+    n_kv_heads: int,
+    position_ids: torch.Tensor,
+):
+    bsz, q_len, d_model = x.shape
+    head_dim = d_model // n_heads
+
+    qkv_out = w.qkv(x)  # shape: (bsz, q_len, (n_heads + 2*n_kv_heads)*head_dim)
+    q_dim = n_heads * head_dim
+    kv_dim = n_kv_heads * head_dim
+
+    q = qkv_out[..., :q_dim].view(bsz, q_len, n_heads, head_dim).transpose(1, 2)
+    k = (
+        qkv_out[..., q_dim : q_dim + kv_dim]
+        .view(bsz, q_len, n_kv_heads, head_dim)
+        .transpose(1, 2)
+    )
+    v = (
+        qkv_out[..., q_dim + kv_dim :]
+        .view(bsz, q_len, n_kv_heads, head_dim)
+        .transpose(1, 2)
+    )
+
+    q = apply_rotary_emb(q, freqs_cis, position_ids, n_heads)
+    k = apply_rotary_emb(k, freqs_cis, position_ids, n_kv_heads)
+
+    if kv_cache is not None:
+        k, v = kv_cache.update(position_ids, k, v)
+
+    out = F.scaled_dot_product_attention(
+        q, k, v, attn_mask=attn_mask, enable_gqa=n_heads != n_kv_heads
+    )
+    out = out.transpose(1, 2).reshape(bsz, q_len, d_model)
+    out = w.proj(out)
+    return out
+
+
+def _attn(
+    x: torch.Tensor,
+    w: torch.Tensor,
+    freqs_cis: torch.Tensor,
+    attn_mask: torch.Tensor,
+    n_heads: int,
+    n_kv_heads: int,
+):
+    bsz, q_len, d_model = x.shape
+    head_dim = d_model // n_heads
+    pos = 0
+
+    qkv_out = w.qkv(x)  # shape: (bsz, q_len, (n_heads + 2*n_kv_heads)*head_dim)
+    q_dim = n_heads * head_dim
+    kv_dim = n_kv_heads * head_dim
+
+    q = qkv_out[..., :q_dim].view(bsz, q_len, n_heads, head_dim).transpose(1, 2)
+    k = (
+        qkv_out[..., q_dim : q_dim + kv_dim]
+        .view(bsz, q_len, n_kv_heads, head_dim)
+        .transpose(1, 2)
+    )
+    v = (
+        qkv_out[..., q_dim + kv_dim :]
+        .view(bsz, q_len, n_kv_heads, head_dim)
+        .transpose(1, 2)
+    )
+
+    position_ids = torch.arange(pos, pos + q_len, dtype=torch.long)
+    q = apply_rotary_emb(q, freqs_cis, position_ids, n_heads)
+    k = apply_rotary_emb(k, freqs_cis, position_ids, n_kv_heads)
+    out = F.scaled_dot_product_attention(
+        q, k, v, attn_mask=attn_mask, enable_gqa=n_heads != n_kv_heads
+    )
+    out = out.transpose(1, 2).reshape(bsz, q_len, d_model)
+    out = w.proj(out)
+    return out
+
+
+def _produce_hidden(inputs_embeds: torch.Tensor, w: nn.Module, config: TextConfig):
+    hidden_BTC = inputs_embeds
+
+    bsz, q_len, d_model = inputs_embeds.shape
+    attn_mask = torch.zeros(q_len, q_len)
+    attn_mask[:730, :730] = 1
+    for i in range(730, q_len):
+        attn_mask[i, : i + 1] = 1
+    attn_mask = attn_mask.to(dtype=torch.bool)
+
+    for i, block in enumerate(w.blocks):
+        l_in = layer_norm(hidden_BTC, block.ln)
+        l_attn = _attn(
+            x=l_in,
+            w=block.attn,
+            freqs_cis=w.freqs_cis,
+            attn_mask=attn_mask,
+            n_heads=config.n_heads,
+            n_kv_heads=config.n_kv_heads,
+        )
+        l_mlp = mlp(l_in, block.mlp)
+        hidden_BTC = hidden_BTC + l_attn + l_mlp
+
+    return hidden_BTC
+
+
+def text_decoder(
+    x: torch.Tensor,
+    w: nn.Module,
+    attn_mask: torch.Tensor,
+    position_ids: torch.Tensor,
+    config: TextConfig,
+):
+    for i, block in enumerate(w.blocks):
+        l_in = layer_norm(x, block.ln)
+        l_attn = attn(
+            l_in,
+            block.attn,
+            freqs_cis=w.freqs_cis,
+            kv_cache=block.kv_cache,
+            attn_mask=attn_mask,
+            n_heads=config.n_heads,
+            n_kv_heads=config.n_kv_heads,
+            position_ids=position_ids,
+        )
+        l_mlp = mlp(l_in, block.mlp)
+        x = x + l_attn + l_mlp
+
+    return x
+
+
+def lm_head(hidden_BTC: torch.Tensor, w: nn.Module):
+    hidden_BC = hidden_BTC[:, -1, :]
+    hidden_BC = layer_norm(hidden_BC, w.post_ln)
+    logits = w.lm_head(hidden_BC)
+    return logits
+
+
+def _lm_head(hidden_BTC: torch.Tensor, w: nn.Module):
+    hidden_BTC = layer_norm(hidden_BTC, w.post_ln)
+    logits = w.lm_head(hidden_BTC)
+    return logits
+
+
+def build_text_model(config: TextConfig, dtype: torch.dtype) -> nn.Module:
+    qkv_dim = int(config.dim * (1 + 2 * config.n_kv_heads / config.n_heads))
+
+    text = nn.ModuleDict(
+        {
+            "blocks": nn.ModuleList(
+                [
+                    nn.ModuleDict(
+                        {
+                            "ln": nn.LayerNorm(config.dim, dtype=dtype),
+                            "attn": nn.ModuleDict(
+                                {
+                                    "qkv": nn.Linear(config.dim, qkv_dim, dtype=dtype),
+                                    "proj": nn.Linear(
+                                        config.dim, config.dim, dtype=dtype
+                                    ),
+                                }
+                            ),
+                            "mlp": nn.ModuleDict(
+                                {
+                                    "fc1": nn.Linear(
+                                        config.dim, config.ff_dim, dtype=dtype
+                                    ),
+                                    "fc2": nn.Linear(
+                                        config.ff_dim, config.dim, dtype=dtype
+                                    ),
+                                }
+                            ),
+                        }
+                    )
+                    for _ in range(config.n_layers)
+                ]
+            ),
+            "post_ln": nn.LayerNorm(config.dim, dtype=dtype),
+            "lm_head": nn.Linear(config.dim, config.vocab_size, dtype=dtype),
+        }
+    )
+    text.wte = nn.Parameter(torch.empty(config.vocab_size, config.dim, dtype=dtype))
+    text.register_buffer(
+        "freqs_cis",
+        precompute_freqs_cis(config.dim // (2 * config.n_heads), config.max_context),
+        persistent=False,
+    )
+
+    return text

utils.py

@@ -0,0 +1,41 @@
+import numpy as np
+
+
+def remove_outlier_points(points_tuples, k_nearest=2, threshold=2.0):
+    """
+    Robust outlier detection for list of (x,y) tuples.
+    Only requires numpy.
+
+    Args:
+        points_tuples: list of (x,y) tuples
+        k_nearest: number of neighbors to consider
+        threshold: multiplier for median distance
+
+    Returns:
+        list: filtered list of (x,y) tuples with outliers removed
+        list: list of booleans indicating which points were kept (True = kept)
+    """
+    points = np.array(points_tuples)
+    n_points = len(points)
+
+    # Calculate pairwise distances manually
+    dist_matrix = np.zeros((n_points, n_points))
+    for i in range(n_points):
+        for j in range(i + 1, n_points):
+            # Euclidean distance between points i and j
+            dist = np.sqrt(np.sum((points[i] - points[j]) ** 2))
+            dist_matrix[i, j] = dist
+            dist_matrix[j, i] = dist
+
+    # Get k nearest neighbors' distances
+    k = min(k_nearest, n_points - 1)
+    neighbor_distances = np.partition(dist_matrix, k, axis=1)[:, :k]
+    avg_neighbor_dist = np.mean(neighbor_distances, axis=1)
+
+    # Calculate mask using median distance
+    median_dist = np.median(avg_neighbor_dist)
+    mask = avg_neighbor_dist <= threshold * median_dist
+
+    # Return filtered tuples and mask
+    filtered_tuples = [t for t, m in zip(points_tuples, mask) if m]
+    return filtered_tuples

vision.py

@@ -0,0 +1,147 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+
+from typing import Union, Tuple
+from PIL import Image
+
+from .layers import attn, layer_norm, linear, mlp
+from .image_crops import overlap_crop_image
+from .config import VisionConfig
+
+if torch.backends.mps.is_available():
+    # Non-divisible input sizes are not implemented on MPS device yet.
+    # https://github.com/pytorch/pytorch/issues/96056
+    def adaptive_avg_pool2d(input, output_size):
+        return F.adaptive_avg_pool2d(input.to("cpu"), output_size).to("mps")
+
+else:
+    adaptive_avg_pool2d = F.adaptive_avg_pool2d
+
+DeviceLike = Union[str, torch.device, int]
+
+
+def prepare_crops(
+    image: Image.Image, config: VisionConfig, device: DeviceLike
+) -> Tuple[torch.Tensor, Tuple[int, int]]:
+    np_image = np.array(image.convert("RGB"))
+    overlap_crops = overlap_crop_image(
+        np_image, max_crops=config.max_crops, overlap_margin=config.overlap_margin
+    )
+    all_crops = overlap_crops["crops"]
+    all_crops = np.transpose(all_crops, (0, 3, 1, 2))
+    all_crops = (
+        torch.from_numpy(all_crops)
+        .to(device=device, dtype=torch.float16)
+        .div_(255.0)
+        .sub_(0.5)
+        .div_(0.5)
+    )
+    return all_crops, overlap_crops["tiling"]
+
+
+def create_patches(x, patch_size):
+    # Original shape: [B, C, H, W]
+    B, C, H, W = x.shape
+    P1 = P2 = patch_size
+
+    # Step 1: Split H and W dimensions into patches
+    # [B, C, H/P1, P1, W/P2, P2]
+    x = x.reshape(B, C, H // P1, P1, W // P2, P2)
+
+    # Step 2: Rearrange dimensions to match target shape
+    # [B, H/P1, W/P2, C, P1, P2]
+    x = x.permute(0, 2, 4, 1, 3, 5)
+
+    # Step 3: Combine dimensions to get final shape
+    # [B, (H/P1)*(W/P2), C*P1*P2]
+    x = x.reshape(B, (H // P1) * (W // P2), C * P1 * P2)
+
+    return x
+
+
+def vision_encoder(input_BCHW: torch.Tensor, w: nn.Module, config: VisionConfig):
+    x = create_patches(input_BCHW, config.enc_patch_size)
+
+    x = linear(x, w.patch_emb)
+    x = x + w.pos_emb
+    for block in w.blocks:
+        x = x + attn(layer_norm(x, block.ln1), block.attn, n_heads=config.enc_n_heads)
+        x = x + mlp(layer_norm(x, block.ln2), block.mlp)
+    x = layer_norm(x, w.post_ln)
+
+    return x
+
+
+def vision_projection(
+    global_features: torch.Tensor,
+    reconstructed: torch.Tensor,
+    w: nn.Module,
+    config: VisionConfig,
+):
+    reconstructed = reconstructed.permute(2, 0, 1)
+    reconstructed = adaptive_avg_pool2d(
+        reconstructed, output_size=(config.enc_n_layers, config.enc_n_layers)
+    )
+    reconstructed = reconstructed.permute(1, 2, 0).view(729, config.enc_dim)
+    final_features = torch.cat([global_features, reconstructed], dim=-1)
+    return mlp(final_features, w.proj_mlp)
+
+
+def build_vision_model(config: VisionConfig, dtype: torch.dtype):
+    patch_dim = config.enc_patch_size * config.enc_patch_size * config.in_channels
+    grid_size = config.crop_size // config.enc_patch_size
+    num_patches = grid_size * grid_size
+
+    vision = nn.ModuleDict(
+        {
+            "patch_emb": nn.Linear(patch_dim, config.enc_dim, dtype=dtype),
+            "blocks": nn.ModuleList(
+                [
+                    nn.ModuleDict(
+                        {
+                            "ln1": nn.LayerNorm(config.enc_dim, dtype=dtype),
+                            "attn": nn.ModuleDict(
+                                {
+                                    "qkv": nn.Linear(
+                                        config.enc_dim, 3 * config.enc_dim, dtype=dtype
+                                    ),
+                                    "proj": nn.Linear(
+                                        config.enc_dim, config.enc_dim, dtype=dtype
+                                    ),
+                                }
+                            ),
+                            "ln2": nn.LayerNorm(config.enc_dim, dtype=dtype),
+                            "mlp": nn.ModuleDict(
+                                {
+                                    "fc1": nn.Linear(
+                                        config.enc_dim, config.enc_ff_dim, dtype=dtype
+                                    ),
+                                    "fc2": nn.Linear(
+                                        config.enc_ff_dim, config.enc_dim, dtype=dtype
+                                    ),
+                                }
+                            ),
+                        }
+                    )
+                    for _ in range(config.enc_n_layers)
+                ]
+            ),
+            "post_ln": nn.LayerNorm(config.enc_dim, dtype=dtype),
+            "proj_mlp": nn.ModuleDict(
+                {
+                    "fc1": nn.Linear(
+                        config.enc_dim * 2, config.proj_inner_dim, dtype=dtype
+                    ),
+                    "fc2": nn.Linear(
+                        config.proj_inner_dim, config.proj_out_dim, dtype=dtype
+                    ),
+                }
+            ),
+        }
+    )
+    vision.pos_emb = nn.Parameter(
+        torch.zeros(1, num_patches, config.enc_dim, dtype=dtype)
+    )
+    return vision