Vyacheslav Pachkov

At this moment there’s no support for building on Windows. The reasons are some parts of Macabuilder use certain posix syscalls such as glob and there are a few bash...

Before moving forward MacaronOS/Macabuilder#5 is required to be implemented

@geohot Thank you! And sure, I would like to continue with this one. Based on my current understanding of tinygrad and Adreno, I would expect to finish in 2-3 weeks....

> How's progress on this? We have a new ETA of completion? > > Sorry to be pushy on this, but we need it for comma. Will up the bounty...

this can now run some tests! ``` ~/tinygrad $ QCOM=1 python test/test_ops.py TestOps.test_tiny_add test_tiny_add (__main__.TestOps.test_tiny_add) ... testing [3, 3] np/tinygrad fp: 0.02 / 685.06 ms bp: nan / nan ms...

this can now run kernels with constants and custom arguments layouts! ``` import tinygrad as tg x = tg.Tensor([1.0, 2.0, 3.0, 4.0, 5.0]) y = x.sin() print(y.numpy()) ``` ``` ~/tinygrad...

> For all these things, a good strategy is to just use the sniffer and copy exactly what it's doing Sure, but this step was done at least a week...

Running gpt2 on qcom! ``` vpachkov@tiny22:~/tinygrad$ QCOM=1 python examples/gpt2.py using QCOM backend using gpt2-medium ram used: 1.42 GB, lm_head.weight : 100%|███████████████████████████████████████████████████████████████████████████████████████████████████████████| 293/293 [00:10

gpt2 on QCOM + HCQ is a little faster than GPU: `ran model in 305.47 ms` vs `ran model in 315.65 ms`

Current status: After excluding tests that fails even with `GPU=1`, running `QCOM=1 python test/test_ops.py` consistently gives the following result: ``` Ran 292 tests in 309.560s FAILED (errors=11, skipped=20) ``` looks...