"""Stage 2 gate 1: the re-authored PyTorch function vs the MLX reference on one real frame. .venv/bin/python check_torch.py [image] [--dtype fp32|fp16] [--device cpu|mps] """ import argparse, json, time import numpy as np, torch from PIL import Image from mlx_vlm import load from d1_common import build_meta, frame, pixels, mlx_logits, readout from questions import QUESTIONS from d1_torch import D1Fixed ap = argparse.ArgumentParser() ap.add_argument("image", nargs="?", default="../screenshot.png") ap.add_argument("--dtype", default="fp32") ap.add_argument("--device", default="cpu") a = ap.parse_args() model, _ = load("LiquidAI/d1-2B-MLX-8bit") eng = model.engine meta = build_meta(eng, QUESTIONS) img = frame(Image.open(a.image)) ref = mlx_logits(eng, img, meta) sys1 = model.system_one(None, QUESTIONS, media=[img])["answers"] assert json.dumps(readout(eng, QUESTIONS, ref, meta), default=float) == json.dumps(sys1, default=float), "readout != system_one" print("host readout of MLX logits == system_one: OK") dt = {"fp32": torch.float32, "fp16": torch.float16}[a.dtype] t = time.time() m = D1Fixed(meta, dt).to(a.device).eval() print("torch model built in %.1fs" % (time.time() - t)) x = torch.from_numpy(pixels(eng, img)).to(a.device) with torch.no_grad(): t = time.time() out = m(x).cpu().numpy() print("torch forward %.1fs" % (time.time() - t)) np.save("out/check_torch_%s_%s.npy" % (a.dtype, a.device), out) err = out - ref psnr = 10 * np.log10(np.abs(ref).max() ** 2 / np.mean(err ** 2)) print("logits PSNR vs MLX-8bit: %.1f dB, max abs err %.3f" % (psnr, np.abs(err).max())) ans = readout(eng, QUESTIONS, out, meta) for k in QUESTIONS: r, o = sys1[k], ans[k] if r["type"] == "choice": print(k, "MLX", r["choice"], round(r["confidence"], 3), "| torch", o["choice"], round(o["confidence"], 3)) else: print(k, "MLX", r["box"], round(r["found"], 3), "| torch", o["box"], round(o["found"], 3))