Executed tinygrad CPU projects¶
Use the guide's pinned setup: Python 3.11+, tinygrad 0.14.0, and Clang for the tested Linux CPU route. All commands below start at the repository root. The short numbered scripts call the readable shared course implementation, so each lesson uses the same environment, dtype and numerical checks.
1. Run the finite progression¶
/tmp/course-tinygrad-venv/bin/python "Phase 5 - Advanced Topics and Specialization/Track E - Autonomous Vehicles/3. tinygrad for Inference/projects/course.py" all --output /tmp/tinygrad-course.json
/tmp/course-tinygrad-venv/bin/python -m unittest discover -s "Phase 5 - Advanced Topics and Specialization/Track E - Autonomous Vehicles/3. tinygrad for Inference/labs" -p "test_*.py" -v
The default makes no dataset/model download. It runs tensor operations, gradients, a tiny synthetic classifier, checkpoint reload, compiler/JIT checks, custom operations and a steering-output contract. The reviewed environment passed all 13 tests: seven mathematical references plus six actual framework tests. Ordinary Python without tinygrad still runs the seven references and explicitly skips the optional framework class.
| Script / lesson | Predict before running | Change after it passes |
|---|---|---|
| 00 intro | Affine/ReLU of [-1,0,2] is [0,1,5] |
Include a negative value whose output clips |
| 01 tensor basics | Transpose [2,3] becomes [3,2] |
Materialize and compare an odd-shaped transpose |
| 02 operation types | Matmul output [[2,11],[-3,6]] |
Change inner dimension to 5 and retain a scalar oracle |
| 03 autograd | Gradient [4,6]; SGD w_20=2(1-.8^20) |
Compare RMSNorm gradients at new smooth values |
| 04 classifier | Initial two-class zero-logit loss is ln(2) |
Add a hidden layer and separate validation data |
| 05 compiler | Realization may change materialization/kernel count | Change input values during JIT replay |
| 06 custom operations | Causal first output equals first value vector | Change only a future token and test earlier outputs |
course.py model_contract |
[1,.1] → -.12 road-wheel radians |
Introduce and reject wrong declared units |
The numbered files support --output. For example, replace course.py all in the first command with 03_autograd.py to run one lesson.
2. Run the explicit real-data extension¶
/tmp/course-tinygrad-venv/bin/python "Phase 5 - Advanced Topics and Specialization/Track E - Autonomous Vehicles/3. tinygrad for Inference/projects/04_mnist.py" --mnist --steps 40 --output /tmp/tinygrad-mnist.json
This downloads/caches MNIST through the pinned loader. It uses 512 train and 256 test images, 32-example batches and a linear classifier; it is not the full dataset benchmark. The reviewed result was loss 2.302599 → 0.502533, test accuracy 202/256, and exact output parity after save/reload. See the recorded run.
No hidden MNIST download occurs in the default synthetic exercise. --steps is bounded to 1–500. Small step counts may intentionally fail the training-progress expectation; such a failure needs interpretation, not a relabeled success.
3. Inspect compilation separately¶
COURSE_TINYGRAD_DEBUG=4 /tmp/course-tinygrad-venv/bin/python "Phase 5 - Advanced Topics and Specialization/Track E - Autonomous Vehicles/3. tinygrad for Inference/projects/05_compiler_pipeline.py" > /tmp/tinygrad-compiler-trace.txt 2>&1
The runner fixes DEV=CPU and BEAM=0. Do not substitute old CLANG=1, DEVICE=CLANG or .schedule() examples from another release. Debug levels and tuning controls are version-specific; inspect the generated trace rather than assuming a level dumps every IR pass.
The recorded CPU report contains three versus one executed model kernels for this tiny expression and four changing-input JIT results. It makes no throughput or latency claim. NOOPT is not a universal switch that disables every mathematical rewrite.
4. Compile a standalone runtime with guarded buffers¶
python3 "Phase 5 - Advanced Topics and Specialization/Track E - Autonomous Vehicles/3. tinygrad for Inference/projects/07_custom_backend.py" --output /tmp/standalone-runtime.json
Project 7 needs Clang and the Python standard library on Linux. It creates a temporary C shared library and binds an explicit function signature. Output elements start as NaNs; guard values surround the result. This catches missing stores and simple tail overwrites while testing sizes 0,1,3,31,33,257 with two inputs each. Recorded 12-case result.
Its filename preserves the original project link, but the implemented milestone is a standalone runtime. There is no fictitious tinygrad device registration and no accelerator performance result. Transfer: add a non-square row-major matmul entry point, explicit dimensions and guarded output; compare with scalar Python before adding timing. A full framework backend is the next separate integration.
5. Submit one evidence bundle¶
Include predictions, test output, package/compiler identity, the JSON report and one failing semantic variation. Explain whether a result was mathematical, actual CPU execution, compiled C execution or an unexecuted extension. A lower loss, fewer kernels, and a valid command are three different observations; connect them with their own checks.