Fixed Tensor Testing

Testing with Fixed Input Tensors

Purpose

Fixed tensor testing ensures deterministic, reproducible results for loss functions and model outputs. Enables verification without training dependencies.

Creating Fixed Tensors

1. Deterministic Seeding

import torch
import numpy as np

# Set all random seeds
torch.manual_seed(42)
np.random.seed(42)
if torch.cuda.is_available():
    torch.cuda.manual_seed_all(42)

2. Creating Test Tensors

# Fixed log probabilities (typically negative)
log_probs = torch.randn(batch_size, seq_len)
# Ensure reasonable log prob range (e.g., -2 to 0)
log_probs = torch.clamp(log_probs, min=-5.0, max=0.0)

# Fixed sequence lengths
seq_lengths = torch.randint(10, 100, (batch_size,))

# Fixed input IDs (for masking)
input_ids = torch.randint(0, vocab_size, (batch_size, seq_len))

3. Example Test Setup

batch_size = 4
seq_len = 10

# Generate fixed tensors
torch.manual_seed(123)
log_probs = torch.randn(batch_size, seq_len)
log_probs = torch.clamp(log_probs, -5.0, 0.0)

# Create mask (e.g., padding)
mask = torch.ones(batch_size, seq_len, dtype=torch.bool)
mask[:, seq_len-2:] = False  # Last 2 tokens are padding

# Lengths accounting for mask
seq_lengths = mask.sum(dim=1).float()

Saving Test Results

Save as .npz (NumPy Compressed)

import numpy as np

results = {
    'losses': loss.cpu().detach().numpy(),
    'log_probs': log_probs.cpu().detach().numpy(),
}

np.savez_compressed('/path/to/results.npz', **results)

Load .npz Files

data = np.load('/path/to/results.npz')
losses = data['losses']
print(f"Shape: {losses.shape}, dtype: {losses.dtype}")

Assertions and Validation

Basic Checks

# Loss should be finite and positive
assert torch.isfinite(loss).all(), "Loss contains NaN or Inf"
assert loss.item() > 0, "Loss should be positive"

# Shape validation
assert loss.shape == expected_shape, f"Shape mismatch: {loss.shape}"

# Range checks
assert loss.item() = -6.0, "Log probs too small"

Reproducibility Checks

# Run computation twice, should get same result
loss1 = compute_loss(fixed_tensors)
loss2 = compute_loss(fixed_tensors)
assert torch.allclose(loss1, loss2), "Loss not reproducible"

Common Test Patterns

1. Unit Test for Loss Function

def test_simpo_loss():
    # Setup fixed inputs
    batch_size = 8
    torch.manual_seed(42)

    log_probs = torch.randn(batch_size, 20)
    seq_lengths = torch.full((batch_size,), 20.0)

    # Compute loss
    beta = 2.0
    gamma = 1.0

    loss = compute_simpo_loss(log_probs, seq_lengths, beta, gamma)

    # Assertions
    assert loss.shape == torch.Size([])
    assert torch.isfinite(loss)
    assert loss.item() > 0

    return loss.item()

2. Gradient Flow Test

# Ensure gradients can backpropagate
x = torch.randn(5, 10, requires_grad=True)
loss = some_loss_function(x)
loss.backward()
assert x.grad is not None
assert not torch.allclose(x.grad, torch.zeros_like(x.grad))

3. Numerical Stability Test

# Test with extreme values
extreme_inputs = [
    torch.full((5,), -100.0),  # Very negative log probs
    torch.full((5,), 0.0),      # Zero log probs
    torch.zeros(5),             # All zeros
]

for inp in extreme_inputs:
    try:
        loss = compute_loss(inp)
        assert torch.isfinite(loss), f"Loss not finite for {inp}"
    except Exception as e:
        print(f"Failed with input {inp}: {e}")

Best Practices

1. Use consistent seeds across test runs
2. Document expected outputs for regression testing
3. Test edge cases (empty batches, single samples, extreme values)
4. Validate shapes before and after operations
5. Check numerical stability with both typical and extreme inputs
6. Save results in standard formats (.npz, .pt, .json)

Debugging Fixed Tensor Tests

# Print intermediate values
def debug_loss(log_probs, seq_lengths, beta, gamma):
    rewards = beta * log_probs / seq_lengths.unsqueeze(1)
    print(f"Rewards shape: {rewards.shape}, min: {rewards.min():.4f}, max: {rewards.max():.4f}")

    # ... rest of computation

    print(f"Final loss: {loss.item():.6f}")
    return loss

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: cxcscmu
• Source: cxcscmu/SkillLearnBench
• License: MIT
• Homepage: https://cxcscmu.github.io/SkillLearnBench

Install and usage instructions live in the source repository linked above.