mirror of
https://github.com/MLSysBook/TinyTorch.git
synced 2026-05-27 08:06:46 -05:00
aa310c7b69
Created systematic tests to verify transformer learning on simple tasks: test_05_transformer_simple_patterns.py: - Test 1: Constant prediction (always predict 5) → 100% ✅ - Test 2: Copy task (failed due to causal masking) → Expected behavior - Test 3: Sequence completion ([0,1,2]→[1,2,3]) → 100% ✅ - Test 4: Pattern repetition ([a,b,a,b,...]) → 100% ✅ test_05_debug_copy_task.py: - Explains why copy task fails (causal masking) - Tests next-token prediction (correct task) → 100% ✅ - Tests memorization vs generalization → 50% (reasonable) Key insight: Autoregressive models predict NEXT token, not SAME token. Position 0 cannot see itself, so "copy" is impossible. The correct task is next-token prediction: [1,2,3,4]→[2,3,4,5] These tests prove the transformer architecture works correctly before attempting full Shakespeare training.
TinyTorch Test Suite
Comprehensive testing organized by purpose and scope.
Test Organization
📦 Module Tests (XX_modulename/)
Purpose: Test individual module functionality
Scope: Single module, isolated behavior
Example: 01_tensor/test_progressive_integration.py
These tests validate that each module works correctly in isolation.
🔗 Integration Tests (integration/)
Purpose: Test cross-module interactions
Scope: Multiple modules working together
Files:
test_gradient_flow.py- CRITICAL: Validates gradients flow through entire training stacktest_end_to_end_training.py- Full training loops (TODO)test_module_compatibility.py- Module interfaces (TODO)
Why this matters:
- Catches bugs that unit tests miss
- Validates the "seams" between modules
- Ensures training actually works end-to-end
🐛 Debugging Tests (debugging/)
Purpose: Catch common student pitfalls
Scope: Pedagogical - teaches debugging
Files:
test_gradient_vanishing.py- Detect/diagnose vanishing gradients (TODO)test_gradient_explosion.py- Detect/diagnose exploding gradients (TODO)test_common_mistakes.py- "Did you forget backward()?" style tests (TODO)
Philosophy: When these tests fail, the error message should teach the student what went wrong and how to fix it.
⚡ Autograd Edge Cases (05_autograd/)
Purpose: Stress-test autograd system
Scope: Autograd internals and edge cases
Files:
test_broadcasting.py- Broadcasting gradient bugs (TODO)test_computation_graph.py- Graph construction edge cases (TODO)test_backward_edge_cases.py- Numerical stability, etc. (TODO)
Running Tests
All tests
pytest tests/ -v
Integration tests only (recommended for debugging training issues)
pytest tests/integration/ -v
Specific test
pytest tests/integration/test_gradient_flow.py -v
Run without pytest
python tests/integration/test_gradient_flow.py
Test Philosophy
- Integration tests catch real bugs: The gradient flow test caught the exact bugs that prevented training
- Descriptive names: Test names should explain what they test
- Good error messages: When tests fail, students should understand why
- Pedagogical value: Tests teach correct usage patterns
Adding New Tests
When adding a test, ask:
- Is it testing one module? → Put in
XX_modulename/ - Is it testing modules working together? → Put in
integration/ - Is it teaching debugging? → Put in
debugging/ - Is it an autograd edge case? → Put in
05_autograd/
Most Important Tests
🔥 Must pass before merging:
integration/test_gradient_flow.py- If this fails, training is broken
📚 Module validation:
- Each module's inline tests (in
modules/source/) - Module-specific tests in
tests/XX_modulename/
Test Coverage Goals
- ✅ All tensor operations have gradient tests
- ✅ All layers compute gradients correctly
- ✅ All activations integrate with autograd
- ✅ All loss functions compute gradients
- ✅ All optimizers update parameters
- ⏳ End-to-end training converges (TODO)
- ⏳ Common pitfalls are detected (TODO)