code/smoke_gpu_batch1.py

"""GPU batch-1 pre-flight for SpatialDiffuseSlot (before real launch / new server).
Loads the EXACT launch config, runs 3 real optimizer steps at batch=1 on ONE GPU:
 - tok_L init OK, forward losses finite, backward+step OK
 - frozen trunk stays frozen (grad None), memory footprint printed
 - EMA-style second forward after step (params updated, still finite)
"""
import torch, time
from omegaconf import OmegaConf
from semanticist.engine.trainer_utils import instantiate_from_config

cfg = OmegaConf.load("configs/tokenizer_l_spatial.yaml")
mp = cfg.trainer.params.model
model = instantiate_from_config(mp).cuda()
model.train()

opt = torch.optim.AdamW([p for p in model.parameters() if p.requires_grad],
                        lr=1e-4, betas=(0.9, 0.95))
print(f"[pre] trainable={sum(p.numel() for p in model.parameters() if p.requires_grad)/1e6:.1f}M "
      f"total={sum(p.numel() for p in model.parameters())/1e6:.1f}M")

losses_log = []
for step in range(3):
    x = torch.randn(1, 3, 256, 256, device="cuda")   # batch=1
    t0 = time.time()
    with torch.autocast("cuda", dtype=torch.bfloat16):
        losses = model(x, sample=False)
        loss = sum(losses.values())
    opt.zero_grad(set_to_none=True)
    loss.backward()
    # frozen trunk check
    bad = [n for n, p in model.dit.named_parameters()
           if n.startswith("blocks.") and p.grad is not None]
    assert not bad, f"frozen trunk got grads: {bad[:3]}"
    torch.nn.utils.clip_grad_norm_(
        [p for p in model.parameters() if p.requires_grad], 3.0)
    opt.step()
    dt = time.time() - t0
    l = {k: float(v) for k, v in losses.items()}
    losses_log.append(l)
    print(f"[step {step}] {l} | {dt:.2f}s | mem {torch.cuda.max_memory_allocated()/2**30:.1f}GiB")
    assert torch.isfinite(loss)

print("GPU BATCH-1 PRE-FLIGHT PASSED ✅ (3 steps, finite, frozen trunk intact)")