HunyuanOCR: document parsing with Transformers

This page is the Transformers contract of the HunyuanOCR MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against tencent/HunyuanOCR. The MIC runner runs the same blocks against each slim fixture and each checkpoint in integration.json. Lines marked # doc-builder: hide only matter there: they swap in the pinned checkpoint or variant and print the result record.

Read a document image

import json  # doc-builder: hide
import os  # doc-builder: hide

import numpy as np
import torch
from transformers import AutoProcessor, HunYuanVLForConditionalGeneration, set_seed

model_id = "tencent/HunyuanOCR"
model_id = os.environ.get("MIC_MODEL", model_id)  # doc-builder: hide
variant = json.loads(os.environ.get("MIC_MODEL_KWARGS", "{}"))  # doc-builder: hide
device = "cuda" if torch.cuda.is_available() else "cpu"
set_seed(0)  # seeds torch, NumPy, and Python: loading, dropout, and data augmentation draw random numbers

processor = AutoProcessor.from_pretrained(model_id)
model = HunYuanVLForConditionalGeneration.from_pretrained(model_id).to(device)
if variant:  # doc-builder: hide
    del model  # doc-builder: hide
    model = HunYuanVLForConditionalGeneration.from_pretrained(model_id, **variant).to(device)  # doc-builder: hide
model.eval()

# Two small synthetic "document" images: white page with a dark bar of text.
def page(bar_y, bar_h):
    img = np.full((128, 128, 3), 255, dtype=np.uint8)
    img[bar_y:bar_y + bar_h, 16:112] = 0
    return img

images = [page(24, 16), page(80, 24)]
prompt = "Extract the text from the image."
messages = [
    [{"role": "user", "content": [{"type": "image", "image": im}, {"type": "text", "text": prompt}]}]
    for im in images
]
texts = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
inputs = processor(text=texts, images=images, return_tensors="pt", padding=True)
inputs = inputs.to(device, dtype=model.dtype)

with torch.no_grad():
    out = model(**inputs)
logits = out.logits
# Left padding: probe the last (real) position of each row, never position 0.
probe = logits[:, -1, :].float()
probe_ids = probe.topk(5, dim=-1).indices
probe_vals = probe.gather(-1, probe_ids)
print(probe_ids.tolist())

One training step

# Language-model labels: copy input_ids and mask padded positions.
labels = inputs["input_ids"].clone()
if "attention_mask" in inputs:
    labels[inputs["attention_mask"] == 0] = -100

model.train()
optimizer = torch.optim.SGD([p for p in model.parameters() if p.requires_grad], lr=0.01)
loss = model(**inputs, labels=labels).loss
loss.backward()
before = {n: p.detach().cpu().clone() for n, p in model.named_parameters() if p.requires_grad}  # doc-builder: hide
optimizer.step()
print(f"loss: {loss.item():.4f}")

if os.environ.get("MIC_RUNNER"):  # doc-builder: hide
    named = dict(model.named_parameters())
    checked = (
        "model.language_model.layers.0.self_attn.q_proj.weight",
        "model.language_model.layers.0.mlp.down_proj.weight",
    )
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "logits": {"probe": probe_vals.tolist()},
        "training": {
            "loss": loss.item(),
            "gradients": {
                n: bool(named[n].grad is not None and torch.count_nonzero(named[n].grad) > 0 and torch.isfinite(named[n].grad).all())
                for n in checked
            },
            "updated": any(not torch.equal(p.detach().cpu(), before[n]) for n, p in model.named_parameters() if n in before),
        },
    }
    print(json.dumps(result))