DeepSeek-R1: causal language modeling with Transformers

This page is the Transformers contract of the DeepSeek-R1 MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against deepseek-ai/DeepSeek-R1. 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.

Predict the next token

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

model_id = "deepseek-ai/DeepSeek-R1"
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 = AutoModelForCausalLM.from_pretrained(model_id).to(device)
if variant:  # doc-builder: hide
    del model  # doc-builder: hide
    model = AutoModelForCausalLM.from_pretrained(model_id, **variant).to(device)  # doc-builder: hide

prompts = [
    "The capital of France is",
    "The capital of Germany is",
]
# Left padding keeps the last position the real next-token position for every row.
tokenizer.padding_side = "left"
inputs = tokenizer(prompts, return_tensors="pt", padding=True).to(device)

model.eval()
with torch.no_grad():
    logits = model(**inputs).logits
next_token_ids = logits[:, -1, :].argmax(dim=-1)
print(tokenizer.batch_decode(next_token_ids[:, None]))

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())
    checked = (
        "model.layers.0.self_attn.q_a_proj.weight",
        "model.layers.0.mlp.gate_proj.weight",
    )
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "next_token_ids": next_token_ids.tolist(),
        "probe_logits": logits[:, -1, :5].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))