MuseGlimmer: image-conditioned text generation with Transformers

This page is the Transformers contract of the MuseGlimmer MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against optimumintel/tiny-random-muse-glimmer. 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 pin the checkpoint or variant and print the result record.

Describe two images

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

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

model_id = "optimum-intel-internal-testing/tiny-random-muse-glimmer"
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 = AutoModelForImageTextToText.from_pretrained(model_id).to(device)
if variant:  # doc-builder: hide
    del model  # doc-builder: hide
    model = AutoModelForImageTextToText.from_pretrained(model_id, **variant).to(device)  # doc-builder: hide

# Two synthetic images that differ in colour, not only in texture.
images = [
    np.full((56, 56, 3), (200, 40, 40), dtype=np.uint8),
    np.full((56, 56, 3), (40, 40, 200), dtype=np.uint8),
]
messages = [
    [
        {"role": "user", "content": [
            {"type": "image", "image": images[0]},
            {"type": "text", "text": "What colour is this image?"},
        ]},
    ],
    [
        {"role": "user", "content": [
            {"type": "image", "image": images[1]},
            {"type": "text", "text": "What colour is this image?"},
        ]},
    ],
]
texts = [
    processor.apply_chat_template(m, add_generation_prompt=True, tokenize=False)
    for m in messages
]
inputs = processor(text=texts, images=images, padding=True, return_tensors="pt")
inputs = inputs.to(device)
if "pixel_values" in inputs:
    inputs["pixel_values"] = inputs["pixel_values"].to(dtype=model.dtype)

model.eval()
with torch.no_grad():
    outputs = model(**inputs)

# Probe the last position of each sequence (left padding puts the real last token at -1).
probe_logits = outputs.logits[:, -1, :]
predicted = probe_logits.argmax(dim=-1)
print(processor.batch_decode(predicted.unsqueeze(-1), skip_special_tokens=True))

One training step

# Causal-LM labels: copy input_ids and mask padded positions.
labels = inputs["input_ids"].clone()
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())  # doc-builder: hide
    checked = (  # doc-builder: hide
        "model.language_model.layers.0.self_attn.q_proj.weight",  # doc-builder: hide
        "model.vision_adapter.fc1.weight",  # doc-builder: hide
    )  # doc-builder: hide
    result = {  # doc-builder: hide
        "runtime": {  # doc-builder: hide
            "device": model.device.type,  # doc-builder: hide
            "dtype": str(model.dtype).removeprefix("torch."),  # doc-builder: hide
            "attn_implementation": model.config._attn_implementation,  # doc-builder: hide
        },  # doc-builder: hide
        "predicted_tokens": predicted.tolist(),  # doc-builder: hide
        "predicted_text": processor.batch_decode(predicted.unsqueeze(-1), skip_special_tokens=True),  # doc-builder: hide
        "answer_logits": probe_logits[:, predicted].float().tolist(),  # doc-builder: hide
        "training": {  # doc-builder: hide
            "loss": loss.item(),  # doc-builder: hide
            "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},  # doc-builder: hide
            "updated": any(not torch.equal(p.detach().cpu(), before[n]) for n, p in model.named_parameters() if n in before),  # doc-builder: hide
        },  # doc-builder: hide
    }  # doc-builder: hide
    print(json.dumps(result))  # doc-builder: hide