rename torchao.testing.float8 to torchao.testing.training

Summary:

Most of the code here is applicable to MX, so making the dir name more
generic.

Test Plan:

```bash
./test/float8/test_everything.sh
```

Reviewers:

Subscribers:

Tasks:

Tags:

ghstack-source-id: e4c8486
ghstack-comment-id: 2991842790
Pull Request resolved: #2415

Author: vkuzo

benchmarks/float8/bench_matmul.py

@@ -16,7 +16,7 @@
     get_name_to_shapes_iter,
 )
 
-from torchao.testing.float8.roofline_utils import get_specs
+from torchao.testing.training.roofline_utils import get_specs
 
 
 def benchmark_fn_in_sec(f, *args, **kwargs):

benchmarks/float8/float8_roofline.py

@@ -63,7 +63,7 @@
 )
 from torchao.prototype.mx_formats import MXLinearConfig
 from torchao.quantization import quantize_
-from torchao.testing.float8.roofline_utils import (
+from torchao.testing.training.roofline_utils import (
     get_float8_mem_sympy,
     get_gemm_time_sympy,
 )

test/float8/test_base.py

@@ -55,7 +55,7 @@
     fp8_tensor_statistics,
     tensor_to_scale,
 )
-from torchao.testing.float8.test_utils import get_test_float8_linear_config
+from torchao.testing.training.test_utils import get_test_float8_linear_config
 from torchao.utils import is_MI300, is_ROCM
 
 random.seed(0)

test/float8/test_compile.py

@@ -37,7 +37,7 @@
     hp_tensor_to_float8_dynamic,
 )
 from torchao.float8.float8_tensor import GemmInputRole, LinearMMConfig, ScaledMMConfig
-from torchao.testing.float8.test_utils import get_test_float8_linear_config
+from torchao.testing.training.test_utils import get_test_float8_linear_config
 
 
 def _test_compile_base(

test/float8/test_dtensor.py

@@ -57,7 +57,7 @@
 )
 from torchao.float8.float8_utils import tensor_to_scale
 from torchao.float8.fsdp_utils import WeightWithDynamicFloat8CastTensor
-from torchao.testing.float8.dtensor_utils import ToyModel
+from torchao.testing.training.dtensor_utils import ToyModel
 
 torch.set_float32_matmul_precision("high")
 

test/float8/test_fsdp2/test_fsdp2.py

@@ -43,7 +43,10 @@
 from torchao.float8.float8_scaling_utils import hp_tensor_to_float8_dynamic
 from torchao.float8.float8_tensor import GemmInputRole
 from torchao.float8.fsdp_utils import WeightWithDynamicFloat8CastTensor
-from torchao.testing.float8.fsdp2_utils import check_parity_bf16_mp, check_parity_no_mp
+from torchao.testing.training.fsdp2_utils import (
+    check_parity_bf16_mp,
+    check_parity_no_mp,
+)
 
 if not is_sm_at_least_89():
     pytest.skip("Unsupported CUDA device capability version", allow_module_level=True)

test/float8/test_fsdp2_tp.py

@@ -32,7 +32,7 @@
     Float8ColwiseParallel,
     Float8RowwiseParallel,
 )
-from torchao.testing.float8.dtensor_utils import ToyModel
+from torchao.testing.training.dtensor_utils import ToyModel
 
 
 def setup_distributed():

test/float8/test_numerics_integration.py

@@ -33,7 +33,7 @@
     convert_to_float8_training,
 )
 from torchao.float8.float8_utils import IS_ROCM, compute_error
-from torchao.testing.float8.test_utils import get_test_float8_linear_config
+from torchao.testing.training.test_utils import get_test_float8_linear_config
 
 torch.manual_seed(0)