clairic-4bit / clairic_runtime /media_cache.py
lucataco's picture
Clairic, MLX 4-bit, with runtime, server and conversion script
1a48d6b verified
Raw History Blame Contribute Delete
3.52 kB
from __future__ import annotations
from collections.abc import Iterable
from typing import TYPE_CHECKING, Any, cast
if TYPE_CHECKING:
from .api import ClairImage
from .documents import ClairDocument
from PIL import Image
from .model import EncodedMedia
import copy
import hashlib
import uuid
from collections import OrderedDict
from dataclasses import dataclass
@dataclass(frozen=True)
class PreparedImages:
images: list[Image.Image]
media: EncodedMedia
payload: list[dict[str, Any]]
salt: str
size: int
source_images: list[Image.Image] | None = None
class ImageCache:
_share_image_row_description: bool
def __init__(
self, max_bytes: int = 128 * 1024 * 1024, max_entries: int = 8
) -> None:
(self.max_bytes, self.max_entries) = (max_bytes, max_entries)
self.namespace = uuid.uuid4().hex
self.entries: OrderedDict[str, PreparedImages] = OrderedDict()
self.size = 0
def clear(self) -> None:
self.entries.clear()
self.size = 0
@staticmethod
def key(images: Iterable[ClairImage | ClairDocument]) -> str:
digest = hashlib.sha256()
for image in images:
content = image.base64.encode()
digest.update(getattr(image, "content_type", "image").encode())
digest.update(str(getattr(image, "rotation", 0)).encode())
digest.update(bytes(["rotation" in getattr(image, "model_fields_set", ())]))
digest.update(len(content).to_bytes(8, "big"))
digest.update(content)
return digest.hexdigest()
def get(self, key: str) -> PreparedImages | None:
entry = self.entries.get(key)
if entry is not None:
self.entries.move_to_end(key)
return entry
def put(
self,
key: str,
images: list[Image.Image],
media: EncodedMedia,
payload: list[dict[str, Any]],
*,
source_images: list[Image.Image] | None = None,
) -> PreparedImages:
tensors = {}
for mapping in (cast(dict[str, Any], media.media), *payload):
for value in mapping.values():
if hasattr(value, "numel"):
tensors[id(value)] = value.numel() * value.element_size()
source_images = images if source_images is None else source_images
retained = {id(image): image for image in [*images, *source_images]}
size = sum(tensors.values()) + sum(
(image.width * image.height * 3 for image in retained.values())
)
if (
media.media
and len(payload) == 1
and ("pixel_values" in media.media)
and ("pixel_values" in payload[0])
):
media = copy.copy(media)
media.media = {
**cast(dict[str, Any], media.media),
"pixel_values": payload[0]["pixel_values"],
}
entry = PreparedImages(
images, media, payload, uuid.uuid4().hex, size, source_images
)
if size > self.max_bytes:
return entry
old = self.entries.pop(key, None)
if old is not None:
self.size -= old.size
while self.entries and (
self.size + size > self.max_bytes or len(self.entries) >= self.max_entries
):
(_, old) = self.entries.popitem(last=False)
self.size -= old.size
self.entries[key] = entry
self.size += size
return entry