Open
Description
For float8, testing needs to be performed on g6.4x machine, hence a new instance is added to the CI/CD pipeline for only float8 tests #587. If we change the current default machine from g5.12x to g6.4x, we'll be having the following benefits:
Cost efficient: We could reduce the CI cost by approx 4x.
g5.12x: $5.672 per hour g5.12x prices
g6.4x: $1.323 per hour g6.4x prices
Unified pipeline: We would reduce the need for adding different machines for different dtypes.
TODO:
- Update default test machine to
amz2023.linux.g6.4xlarge.experimental.nvidia.gpu
- Update tests to support the new machine:
- test/sparsity/test_structured_sparsifier.py::TestBaseStructuredSparsifier::test_prune_linear_linear
- test/sparsity/test_structured_sparsifier.py::TestBaseStructuredSparsifier::test_prune_linear_bias_linear
- test/sparsity/test_structured_sparsifier.py::TestBaseStructuredSparsifier::test_prune_linear_activation_linear
- test/prototype/mx_formats/test_mx_tensor.py::test_to_mx_from_mx_compile_numeric
- test/prototype/test_low_bit_optim.py::TestFSDP2::test_fsdp2
- test/integration/test_integration.py::TestSubclass::test_int8_weight_only_quant_with_freeze
- test/dtypes/test_nf4.py::TestQLoRA::test_qlora_fsdp2