"""A torch-free stand-in for the Transformers `Qwen3VLProcessor` used by the Clairic request pipeline. Not part of the Cloudflare release. The upstream pipeline calls four things on its processor: the chat template, `replace_image_token`, `get_text_with_replacements`, and `processor(text=..., images=...)`. This class provides exactly those on top of the tokenizer and mlx-vlm's NumPy image processor, so Transformers' torchvision-based image and video processors are not needed. """ from __future__ import annotations import json from pathlib import Path from typing import Any import numpy as np MIN_PIXELS = 65_536 class ClairicProcessor: image_token = "<|image_pad|>" def __init__(self, path: Path, tokenizer: Any, max_pixels: int) -> None: from mlx_vlm.models.qwen3_vl.processing_qwen3_vl import Qwen3VLImageProcessor config = json.loads((path / "preprocessor_config.json").read_text()) self.tokenizer = tokenizer self.image_token_id = tokenizer.convert_tokens_to_ids(self.image_token) self.image_processor = Qwen3VLImageProcessor( patch_size=config["patch_size"], temporal_patch_size=config["temporal_patch_size"], merge_size=config["merge_size"], min_pixels=config["size"].get("shortest_edge", MIN_PIXELS), max_pixels=max_pixels, image_mean=config["image_mean"], image_std=config["image_std"]) def apply_chat_template(self, messages: list[dict[str, Any]], **kwargs: Any) -> Any: return self.tokenizer.apply_chat_template(messages, **kwargs) def replace_image_token(self, image_inputs: dict[str, Any], image_idx: int) -> str: merge = self.image_processor.merge_size ** 2 return self.image_token * (int(np.prod(np.asarray(image_inputs["image_grid_thw"])[image_idx])) // merge) def get_text_with_replacements(self, text: list[str], images_replacements: list[str] = (), videos_replacements: list[str] = (), audio_replacements: list[str] = ()) -> tuple[list[str], list]: """Replace each image placeholder, in order across the batch, with its expanded run of placeholders.""" if videos_replacements or audio_replacements: raise ValueError("Clairic accepts images only") replacements = iter(images_replacements) expanded = [] for sample in text: pieces = sample.split(self.image_token) if len(images_replacements) == 0: expanded.append(sample) continue expanded.append("".join(piece + (next(replacements) if index < len(pieces) - 1 else "") for index, piece in enumerate(pieces))) return expanded, [] def __call__(self, text: list[str], images: list[Any], return_tensors: str | None = None) -> dict[str, Any]: processed = self.image_processor(list(images)) media = {"pixel_values": np.asarray(processed["pixel_values"], dtype=np.float32), "image_grid_thw": np.asarray(processed["image_grid_thw"], dtype=np.int64)} expanded, _ = self.get_text_with_replacements( list(text), [self.replace_image_token(media, index) for index in range(len(images))]) ids = np.asarray(self.tokenizer(expanded, add_special_tokens=False)["input_ids"], dtype=np.int64) return {"input_ids": ids, "attention_mask": np.ones_like(ids), **media, "mm_token_type_ids": (ids == self.image_token_id).astype(np.int64)}