# SwitchTransformers: text-to-text generation with Transformers

This page is the Transformers contract of the SwitchTransformers MIC and also a usage example. [doc-builder](https://github.com/huggingface/doc-builder/blob/main/docs/runnable-code-blocks.md) renders the `runnable` blocks below as documentation, and its pytest plugin runs them as written against `google/switch-base-8`. The MIC runner runs the same blocks against the slim fixture and the checkpoint in [`integration.json`](../../integration.json). Lines marked `# doc-builder: hide` only matter there: they swap in the pinned checkpoint or variant and print the result record.

## Generate a batch

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

model_id = "google/switch-base-8"
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

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

sentences = [
    "A <extra_id_0> walks into a bar and orders a <extra_id_1> with a pinch of <extra_id_2>.",
    "The <extra_id_0> is the best place to <extra_id_1> in town.",
]
inputs = tokenizer(sentences, return_tensors="pt", padding=True).to(device)
generated = model.generate(**inputs, max_new_tokens=12, do_sample=False)
print(tokenizer.batch_decode(generated, skip_special_tokens=True))
```

## Decode a batch of token ids

`decode` accepts a batch and returns one string per sequence, cleaning up the spaces the tokenizer puts before punctuation.

```py runnable:generate:3
punctuated = ["Yes , it is .", "Hello , world ! How are you ?"]
roundtrip = tokenizer.decode(tokenizer(punctuated).input_ids, skip_special_tokens=True)
print(roundtrip)
```

## One training step

```py runnable:generate:3
targets = [
    "A man walks into a bar and orders a beer with a pinch of salt.",
    "The park is the best place to relax in town.",
]
labels = tokenizer(text_target=targets, return_tensors="pt", padding=True).input_ids.to(device)
labels[labels == tokenizer.pad_token_id] = -100

with torch.no_grad():
    # Teacher-forced log-probabilities of the reference translations.
    logprobs = model(**inputs, labels=labels).logits.float().log_softmax(-1)
    label_logprobs = logprobs.gather(-1, labels.clamp(min=0)[..., None])[..., 0].masked_fill(labels == -100, 0.0)

model.train()
optimizer = torch.optim.SGD(model.parameters(), lr=0.01)
loss = model(**inputs, labels=labels).loss
loss.backward()
before = {name: p.detach().cpu().clone() for name, 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 = (
        "encoder.block.0.layer.0.SelfAttention.q.weight",
        "decoder.block.0.layer.1.EncDecAttention.q.weight",
        "encoder.block.1.layer.1.mlp.experts.expert_0.wi.weight",
        "decoder.block.1.layer.2.mlp.experts.expert_0.wi.weight",
    )
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "generated_tokens": generated.tolist(),
        "generated_text": tokenizer.batch_decode(generated, skip_special_tokens=True),
        "label_logprobs": label_logprobs.tolist(),
        "decoded_roundtrip": roundtrip,
        "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))
```