{ai}[foss/2023a] tinygrad v0.8.0 w/ CUDA 12.1.1

[VRehnberg]

(created using eb --new-pr)
requires:

• {lib}[GCCcore/12.3.0] PyTorch-bundle v2.1.2, SoX v14.4.2, libmad v0.15.1b, ... w/ CUDA 12.1.1 #19987
• {lib,math}[GCCcore/12.3.0,gfbf/2023a] Transformers v4.39.3, tokenizers v0.15.2 #20322
• {lib}[GCCcore/12.3.0] tiktoken v0.6.0 #20336

[VRehnberg]

Test report by @VRehnberg
FAILED
Build succeeded for 0 out of 1 (1 easyconfigs in total)
alvis1-03 - Linux Rocky Linux 8.9, x86_64, Intel(R) Xeon(R) Gold 6244 CPU @ 3.60GHz, 1 x NVIDIA Tesla V100-SXM2-32GB, 550.54.14, Python 3.6.8
See https://gist.github.com/VRehnberg/f19c66d9244dc4fedc21af02b3694a56 for a full test report.

[VRehnberg]

FAILED test/test_dtype.py::TestHalfDtype::test_casts_from - RuntimeError: compile failed: (14): error: more than one conversion function from "half" to "char" applies:
FAILED test/test_dtype.py::TestHalfDtype::test_casts_to - RuntimeError: compile failed: (14): error: more than one constructor applies to convert from "long" to "__half":
FAILED test/test_dtype.py::TestHalfDtype::test_upcast_ops - RuntimeError: compile failed: (15): error: more than one constructor applies to convert from "long" to "__half":
FAILED test/test_dtype.py::TestInt8Dtype::test_casts_to - RuntimeError: compile failed: (14): error: more than one conversion function from "half" to "char" applies:
FAILED test/test_dtype.py::TestInt8Dtype::test_upcast_ops - RuntimeError: compile failed: (18): error: more than one constructor applies to convert from "unsigned long" to "__half":
FAILED test/test_dtype.py::TestUint8Dtype::test_casts_to - RuntimeError: compile failed: (14): error: more than one conversion function from "half" to "unsigned char" applies:
FAILED test/test_dtype.py::TestInt16Dtype::test_upcast_ops - RuntimeError: compile failed: (18): error: more than one constructor applies to convert from "unsigned long" to "__half":
FAILED test/test_dtype.py::TestInt32Dtype::test_upcast_ops - RuntimeError: compile failed: (18): error: more than one constructor applies to convert from "unsigned long" to "__half":
FAILED test/test_dtype.py::TestInt64Dtype::test_casts_from - RuntimeError: compile failed: (14): error: more than one constructor applies to convert from "long" to "__half":
FAILED test/test_dtype.py::TestInt64Dtype::test_casts_to - RuntimeError: compile failed: (14): error: more than one conversion function from "half" to "long" applies:
FAILED test/test_dtype.py::TestInt64Dtype::test_same_size_ops - RuntimeError: compile failed: (18): error: more than one constructor applies to convert from "long" to "__half":
FAILED test/test_dtype.py::TestInt64Dtype::test_upcast_to_ops - RuntimeError: compile failed: (15): error: more than one constructor applies to convert from "long" to "__half":
FAILED test/test_dtype.py::TestUint64Dtype::test_casts_from - RuntimeError: compile failed: (14): error: more than one constructor applies to convert from "unsigned long" to "__half":
FAILED test/test_dtype.py::TestUint64Dtype::test_casts_to - RuntimeError: compile failed: (14): error: more than one conversion function from "half" to "unsigned long" applies:
FAILED test/test_dtype.py::TestUint64Dtype::test_same_size_ops - RuntimeError: compile failed: (18): error: more than one constructor applies to convert from "unsigned long" to "__half":
FAILED test/test_dtype.py::TestUint64Dtype::test_upcast_to_ops - RuntimeError: compile failed: (18): error: more than one constructor applies to convert from "unsigned long" to "__half":
FAILED test/test_dtype_alu.py::TestDTypeALU::test_float_midcast_int32 - AssertionError:
FAILED test/test_linearizer_failures.py::TestLinearizerFailures::test_failure_8 - AssertionError: assert 'COMPILE_ERROR' == 'PASS'
FAILED test/extra/test_export_model.py::TextModelExport::test_multi_output_model_export - AttributeError: 'CUDAGraph' object has no attribute 'prg'
FAILED test/models/test_real_world.py::TestRealWorld::test_gpt2 - AssertionError: only 8 out of 468 were jitted
FAILED test/models/test_real_world.py::TestRealWorld::test_llama - AssertionError: only 12 out of 719 were jitted

I don't know what error: more than one constructor applies to convert from ... means. Before I patched the error was that cuda_fp16.h couldn't be found.

[VRehnberg]

I believe a pip install in a virtualenv --system-site-packages didn't have this issue so I suppose I just have to diff the results until I find something I can patch.

[VRehnberg]

I believe a pip install in a virtualenv --system-site-packages didn't have this issue so I suppose I just have to diff the results until I find something I can patch.

I've been able to replicate the issue with pip installed tinygrad. Not sure what was different before.

[VRehnberg]

Forcing Clang CLANG=1 python test/test_dtype.py solves this issue (but I suspect it will have performance drawbacks). Micket helped me by pointing out that this the errors probably are due to long being underdetermined (and probably same issue for char then). I haven't found where long and char are specified.

[VRehnberg]

So here is where the dtypes are defined https://github.com/tinygrad/tinygrad/blob/v0.8.0/tinygrad/dtype.py#L51-L64 now the remaining issue what to replace char with. There's no int8_t to replace it with. unsigned char -> uint8_t and long -> long long seems to have gone fine.

[VRehnberg]

Test report by @VRehnberg
FAILED
Build succeeded for 0 out of 1 (1 easyconfigs in total)
alvis1-05 - Linux Rocky Linux 8.9, x86_64, Intel(R) Xeon(R) Gold 6244 CPU @ 3.60GHz, 1 x NVIDIA Tesla V100-SXM2-32GB, 550.54.14, Python 3.6.8
See https://gist.github.com/VRehnberg/59824f85ce3d224c10edde5196b19b7f for a full test report.