"""Check the installed environment on the GPU: OpenCV, PyTorch and PaddlePaddle on one GPU in one process, the engine module, and one OCR call. Exits with a message if any of that fails. uv run --frozen python selfcheck.py""" import sys from pathlib import Path sys.path.insert(0, str(Path(__file__).resolve().parent)) import ctypes.util, json, sys from pathlib import Path def fail(msg): print("selfcheck: FAILED: " + msg, file=sys.stderr); sys.exit(1) try: import cv2 except ImportError as e: fail(f"OpenCV does not import ({e}); install the system packages libgl1 and libglib2.0-0") try: import torch except ImportError as e: fail(f"torch does not import ({e}); PyTorch's CUDA libraries are not the ones installed; run `uv sync --frozen` again") if not torch.cuda.is_available(): fail("torch sees no GPU") x = torch.randn(4, 3, 64, 64, device="cuda:0", dtype=torch.bfloat16) y = torch.nn.functional.conv2d(x, torch.randn(8, 3, 3, 3, device="cuda:0", dtype=torch.bfloat16)); torch.cuda.synchronize() if not bool(torch.isfinite(y.float()).all()): fail("a convolution on the GPU gave non-finite values") import paddle if not paddle.device.is_compiled_with_cuda(): fail("this PaddlePaddle build has no GPU support") import peshgi as engine_module # the engine, without loading the model import numpy as np from huggingface_hub import snapshot_download from paddleocr import PaddleOCR from PIL import Image, ImageDraw, ImageFont oc = json.loads((Path(engine_module.__file__).parent / "peshgi.json").read_text())["ocr"] det = snapshot_download(oc["detection"]["repo"], revision=oc["detection"]["revision"]) rec = snapshot_download(oc["recognition"]["repo"], revision=oc["recognition"]["revision"]) try: ocr = PaddleOCR(text_detection_model_name=oc["detection"]["name"], text_detection_model_dir=str(det), text_recognition_model_name=oc["recognition"]["name"], text_recognition_model_dir=str(rec), use_doc_orientation_classify=False, use_doc_unwarping=False, use_textline_orientation=False, text_det_limit_side_len=oc["det_limit_side_len"], text_det_limit_type="max", text_recognition_batch_size=oc["rec_batch_size"], device="gpu:0") im = Image.new("RGB", (640, 160), (255, 255, 255)) try: font = ImageFont.load_default(size=48) except TypeError: font = ImageFont.load_default() ImageDraw.Draw(im).text((30, 50), "INSTALL 2468", fill=(0, 0, 0), font=font) texts = [] for r in ocr.predict(np.ascontiguousarray(np.asarray(im)[:, :, ::-1])): j = r.json["res"] if hasattr(r, "json") else r texts += list(j.get("rec_texts", [])) except Exception as e: fail(f"the OCR model did not run on the GPU ({type(e).__name__}: {e})") if "2468" not in " ".join(texts): fail(f"the OCR model read {texts!r} instead of 'INSTALL 2468'") y = torch.nn.functional.conv2d(x, torch.randn(8, 3, 3, 3, device="cuda:0", dtype=torch.bfloat16)); torch.cuda.synchronize() print(f"selfcheck: passed (torch {torch.__version__}, paddle {paddle.__version__}, {torch.cuda.get_device_name(0)}, OCR read {texts})")