GPTBigCode: causal language modeling with Transformers

This page is the Transformers contract of the GPTBigCode MIC and also a usage example. doc-builder renders the runnable blocks below as documentation, and its pytest plugin runs them as written against bigcode/tiny_starcoder_py. 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 = "bigcode/tiny_starcoder_py"
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)
# This checkpoint defines no pad token; reuse the end-of-text token and pad on the left so the last position is a real token.
tokenizer.pad_token = tokenizer.eos_token
tokenizer.padding_side = "left"
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 = [
    "def add(a, b):\n    return a + b\n\ndef sub(a, b):\n    return a - b\n\ndef mul(a, b):\n    return",
    "import numpy as np\n\ndef mean(x):\n    return np.",
]
inputs = tokenizer(prompts, return_tensors="pt", padding=True).to(device)

with torch.no_grad():
    logits = model(**inputs).logits
# Left padding: the last position is the final real token of each prompt.
next_logits = logits[:, -1, :]
predicted = next_token_logits = next_token_logits if False else next_logits.argmax(dim=-1)  # doc-builder: hide
predicted = next_logits.argmax(dim=-1)
print(tokenizer.convert_ids_to_tokens(predicted))

One training step

# Language-model labels: copy the input ids and mask padded positions.
labels = inputs["input_ids"].clone()
labels[labels == tokenizer.pad_token_id] = -100

model.train()
optimizer = torch.optim.SGD(model.parameters(), 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 = ("transformer.h.0.attn.c_attn.weight", "transformer.h.0.mlp.c_fc.weight")
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "predicted_tokens": predicted.tolist(),
        "predicted_text": tokenizer.convert_ids_to_tokens(predicted),
        "next_token_logits": next_logits.gather(1, predicted.unsqueeze(-1)).squeeze(-1).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))