tarekziade's picture
tarekziade HF Staff
Publish MIC usage, slim fixture and reviewed Transformers expectations
64c7620 verified
|
Raw History Blame Contribute Delete
4.56 kB

GLM-5.3-Flash: image + text generation with Transformers

This page is the Transformers contract of the GLM-5.3-Flash MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against zai-org/GLM-5.3-Flash. The MIC runner runs the same blocks against the slim fixture and the 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.

Describe two images

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

import torch
from transformers import AutoProcessor, Glm5NextForConditionalGeneration, set_seed

model_id = "zai-org/GLM-5.3-Flash"
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 = Glm5NextForConditionalGeneration.from_pretrained(model_id).to(device)
if variant:  # doc-builder: hide
    del model  # doc-builder: hide
    model = Glm5NextForConditionalGeneration.from_pretrained(model_id, **variant).to(device)  # doc-builder: hide

# Two distinct synthetic images so the batch is meaningful; replace with real images in your app.
images = [torch.randint(0, 256, (3, 56, 56), dtype=torch.uint8) for _ in range(2)]
prompts = [
    "Describe this image in one sentence.",
    "What colors are dominant in this image?",
]
messages = [
    [{"role": "user", "content": [{"type": "image"}, {"type": "text", "text": prompt}]}]
    for prompt in prompts
]
# Render the chat text first, then let the processor pair it with the 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)

model.eval()
with torch.no_grad():
    logits = model(**inputs).logits
# With right padding, the last real token of each row is the next-token probe.
last = inputs["attention_mask"].sum(dim=-1) - 1
probe = logits[torch.arange(logits.shape[0]), last]
predicted = probe.argmax(dim=-1)
print(processor.tokenizer.convert_ids_to_tokens(predicted.tolist()))

One training step

# Language-model 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())
    checked = (
        "model.language_model.layers.0.self_attn.q_proj.weight",
        "model.language_model.layers.0.mlp.gate_proj.weight",
        "model.visual.blocks.0.attn.qkv.weight",
    )
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
            "attn_implementations": {n: getattr(model.config, n)._attn_implementation for n in sorted(model.config.sub_configs) if getattr(getattr(model.config, n, None), '_attn_implementation', None)},
        },
        "predicted_tokens": processor.tokenizer.convert_ids_to_tokens(predicted.tolist()),
        "answer_logits": probe[:, predicted].float().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))