qazal
qazal
status: This is in a good spot now - time to stress test multi output
20 of the possible toposorts for `test.models.test_mnist.TestMNIST.test_adam_onestep` - each color is a LazyBuffer https://tiny-tools-client.vercel.app/?id=c44a8348fff042c5801ff899c4f39913
the sort and kernel runs are generally fast, validation is slow: ```py # assert all LazyBuffers realized correctly for lb, bufs in outputs.items(): ground_truth = np.frombuffer(bufs[0].as_buffer(), bufs[0].dtype.np) for buf in...
This diff creates new Buffers for each permutation - makes fuzzing multiple toposorts in parallel possible.
`test/models/test_train.py::TestTrain::test_convnext` This is the graph: https://github.com/tinygrad/tinygrad/assets/77887910/c99105af-52c3-4e40-8035-e7db9cd789cf Oh it's actually because it keeps recursing the realized child's children - it should break. For this reduce: `(2, 56, 56, 64, 1)` ...
#4180 prereq for green CI
the test failed because it was fusing 3 kernels into 1 xd fyi @chaosagent : ```py # shared between two outputs shared = x.sum().half().float() a = shared * 2 b...
4180 does the job
This diff is very hard to review, can you break it down to isolated PRs?
@0xtimmy can you provide a status update? What are the blockers to getting the linearizer changes merged?