MiniCPM3: causal language modeling with Transformers

This page is the Transformers contract of the MiniCPM3 MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against openbmb/MiniCPM3-4B. 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.

Inference: next-token logits

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

model_id = "openbmb/MiniCPM3-4B"
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)
tokenizer.pad_token = tokenizer.eos_token  # the checkpoint defines no padding token
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

# Two distinct prompts so batch perturbation is meaningful
prompts = ["The capital of France is", "The largest planet in our solar system is"]
inputs = tokenizer(prompts, return_tensors="pt", padding=True).to(device)

model.eval()
with torch.no_grad():
    outputs = model(**inputs)
logits = outputs.logits  # (batch, seq_len, vocab_size)

# Report a few selected logits at the last real (non-padding) token of each sequence
last = inputs["attention_mask"].sum(dim=-1) - 1
last_logits = logits[torch.arange(logits.shape[0], device=device), last].float()
probe = torch.tensor([0, 100, 5000, 73000], device=device)
selected = last_logits[:, probe]
print(selected)

One training step

# Freeze everything except the first layer's MLP gate and the final norm
for param in model.parameters():
    param.requires_grad = False
model.model.layers[0].mlp.gate_proj.weight.requires_grad = True
model.model.norm.weight.requires_grad = True

# Labels: copy input_ids, set padded positions to -100 (the model shifts internally)
input_ids = inputs["input_ids"]
labels = input_ids.clone()
if "attention_mask" in inputs:
    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.mlp.gate_proj.weight", "model.norm.weight")
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "logits": selected.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))