EmbeddingGemma: multimodal embeddings

This page is the MIC integration contract for google/embeddinggemma-2. It shows the public transformers API for computing text and image embeddings and for running a short training step on the model's own parameters. The runnable blocks below are rendered by doc-builder and executed by the MIC runner against the slim fixture and the pinned checkpoint. Lines marked # doc-builder: hide are runner plumbing and are not shown in the rendered documentation.

Compute text and image embeddings

import torch
from transformers import AutoModel, AutoProcessor, set_seed
import json  # doc-builder: hide
import os  # doc-builder: hide

model_id = "google/embeddinggemma-2"
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)  # reproducible synthetic inputs and dropout

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

texts = [
    "What causes the northern lights?",
    "Charged particles from the sun cause the northern lights.",
]
# Two small synthetic RGB images; the processor resizes and normalizes them.
images = [torch.rand(3, 224, 224) for _ in range(2)]

text_inputs = processor(text=texts, return_tensors="pt", padding=True)
text_inputs = text_inputs.to(device)
# The processor packs a list of images into one sequence: embed each on its own.
image_inputs = [processor(images=[image], return_tensors="pt").to(device, dtype=model.dtype) for image in images]

with torch.no_grad():
    text_out = model(**text_inputs)
    image_outs = [model(**inputs) for inputs in image_inputs]

# EmbeddingGemma2Model returns hidden states; mean-pool them into embeddings.
def mean_pool(last_hidden_state, attention_mask):
    mask = attention_mask.unsqueeze(-1).to(last_hidden_state.dtype)
    return (last_hidden_state * mask).sum(dim=1) / mask.sum(dim=1).clamp(min=1e-9)

text_emb = mean_pool(text_out.last_hidden_state, text_inputs["attention_mask"])
image_emb = torch.cat([out.last_hidden_state.mean(dim=1) for out in image_outs])  # image tokens are never padded
similarity = torch.nn.functional.cosine_similarity(text_emb[0], text_emb[1], dim=0)
print(f"similarity: {similarity.item():.4f}")

One training step

# In-batch contrastive step: each text should match its own image. The embeddings
# are recomputed with gradients (the ones above were taken under no_grad).
model.train()
optimizer = torch.optim.SGD(model.parameters(), lr=1e-4)

text_train = mean_pool(model(**text_inputs).last_hidden_state, text_inputs["attention_mask"])
image_train = torch.cat([model(**inputs).last_hidden_state.mean(dim=1) for inputs in image_inputs])
scores = torch.nn.functional.normalize(text_train.float(), dim=-1) @ torch.nn.functional.normalize(image_train.float(), dim=-1).T
loss = torch.nn.functional.cross_entropy(scores / 0.05, torch.arange(len(scores), device=device))
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 = (
        "language_model.layers.0.self_attn.q_proj.weight",
        "language_model.layers.0.mlp.down_proj.weight",
        "vision_tower.encoder.layers.0.self_attn.q_proj.linear.weight",
    )
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "text_embeddings": text_emb.float().tolist(),
        "image_embeddings": image_emb.float().tolist(),
        "similarity": similarity.item(),
        "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))