Inkling: multimodal conditional generation with Transformers

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

Generate a continuation

import torch
from transformers import AutoProcessor, InklingForConditionalGeneration, set_seed
import json  # doc-builder: hide
import os  # doc-builder: hide

model_id = "thinkingmachines/Inkling"
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

processor = AutoProcessor.from_pretrained(model_id)
model = InklingForConditionalGeneration.from_pretrained(model_id).to(device)
if variant:  # doc-builder: hide
    del model  # doc-builder: hide
    model = InklingForConditionalGeneration.from_pretrained(model_id, **variant).to(device)  # doc-builder: hide
model.eval()

# The checkpoint ships no pad token; reuse the EOS token so a two-prompt batch can be padded.
if processor.tokenizer.pad_token is None:
    # Inkling's tokenizer defines no pad or eos token; pad with its end-of-text token.
    processor.tokenizer.pad_token = processor.tokenizer.eos_token or "<|endoftext|>"

# Two distinct prompts so a batch perturbation is meaningful; right padding keeps the last position real.
prompts = ["Describe the scene in one sentence.", "What sound is in the clip?"]
inputs = processor(text=prompts, return_tensors="pt", padding=True)
inputs = inputs.to(device)

with torch.no_grad():
    generated = model.generate(**inputs, max_new_tokens=8)
print(processor.batch_decode(generated, skip_special_tokens=True))

One training step

# Language-model labels: copy input_ids and mask padded positions with -100.
labels = inputs["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.language_model.layers.0.self_attn.q_proj.weight",
        "model.language_model.layers.0.mlp.up_proj.weight",
        "lm_head.weight",
    )
    probe = inputs["input_ids"].shape[-1] - 1
    result = {
        "runtime": {
            "device": model.device.type,
            "dtype": str(model.dtype).removeprefix("torch."),
            "attn_implementation": model.config._attn_implementation,
        },
        "generated_ids": generated.tolist(),
        "logits_probe": model(**inputs).logits[:, probe, :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))