BART: encoder-decoder forward with Transformers

This page is the Transformers contract of the BART MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against facebook/bart-base. 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.

Run the encoder-decoder

import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, set_seed
import json  # doc-builder: hide
import os  # doc-builder: hide

model_id = "facebook/bart-base"
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 and dropout draw random numbers

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForSeq2SeqLM.from_pretrained(model_id).to(device)
if variant:  # doc-builder: hide
    del model  # doc-builder: hide
    model = AutoModelForSeq2SeqLM.from_pretrained(model_id, **variant).to(device)  # doc-builder: hide
model.eval()

sentences = [
    "The quick brown fox jumps over the lazy dog.",
    "Transformers are useful for many natural language tasks.",
]
inputs = tokenizer(sentences, return_tensors="pt", padding=True).to(device)
# BART is an encoder-decoder: give the decoder a short target prefix.
decoder_input_ids = torch.full((len(sentences), 1), model.config.decoder_start_token_id, dtype=torch.long, device=device)

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

encoder_shape = list(outputs.encoder_last_hidden_state.shape)
# Probe a few decoder logits rather than the whole vocabulary.
probe = outputs.logits[0, 0, :5].float().tolist()
print(encoder_shape)
print(probe)

One training step

# Teacher forcing: the model shifts the labels right to build the decoder inputs.
labels = inputs["input_ids"].clone()
labels[labels == tokenizer.pad_token_id] = -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.encoder.layers.0.self_attn.q_proj.weight",
        "model.decoder.layers.0.self_attn.q_proj.weight",
        "model.shared.weight",
    )
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "encoder_last_hidden_shape": encoder_shape,
        "decoder_logits_probe": probe,
        "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))