SkillCompiler/data/skills-bench/tasks-extra/mhc-layer-impl/environment/skills/modal-gpu/references/results.md

Returning Results from Modal

Basic Return Values

Modal functions can return Python objects (dicts, lists, etc.):

@app.function(gpu="A100", image=image, timeout=3600)
def train():
    # ... training code ...

    return {
        "final_loss": float(metrics["final_val_loss"]),
        "grad_norm_std": float(metrics["grad_norm_std"]),
        "max_grad_norm": float(metrics["max_grad_norm"]),
        "training_steps": int(metrics["steps"]),
    }

@app.local_entrypoint()
def main():
    results = train.remote()
    print(f"Results: {results}")

    # Save locally
    import json
    with open("results.json", "w") as f:
        json.dump(results, f, indent=2)

Supported Return Types

Type	Support	Notes
dict	Yes	Most common for metrics
list	Yes	For sequences of results
tuple	Yes	For multiple return values
str, int, float	Yes	Primitive types
numpy array	Limited	Convert to list first
torch tensor	No	Convert to Python types

Converting Tensors

Always convert tensors to Python types before returning:

@app.function(gpu="A100", image=image, timeout=3600)
def train():
    import torch

    # ... training ...

    # WRONG - can't return tensors
    # return {"loss": loss_tensor}

    # CORRECT - convert to float
    return {
        "loss": float(loss_tensor.item()),
        "losses": [float(l) for l in loss_history],
    }

Complete Example

import modal

app = modal.App("model-training")

image = modal.Image.debian_slim(python_version="3.11").pip_install(
    "torch",
    "einops",
    "numpy",
    "huggingface_hub",
)

@app.function(gpu="A100", image=image, timeout=3600)
def train_model():
    # ... training code ...

    return {
        "final_loss": metrics["final_val_loss"],
        "training_steps": total_steps,
        "gpu": torch.cuda.get_device_name(0),
    }

@app.local_entrypoint()
def main():
    results = train_model.remote()
    print(f"\nResults: {results}")

    import json
    with open("results.json", "w") as f:
        json.dump(results, f, indent=2)

Handling Failures

If a function fails, no result is returned. Use try/except to handle errors:

@app.function(gpu="A100", image=image, timeout=3600)
def train():
    try:
        # ... training code ...
        return {"status": "success", "loss": final_loss}
    except Exception as e:
        return {"status": "error", "message": str(e)}