OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

Revert "Add test coverage to tf32 in max autotune mm configs (#161545)"

Browse Source 
This reverts commit e9d34b2438.

Reverted https://github.com/pytorch/pytorch/pull/161545 on behalf of https://github.com/atalman due to inductor/test_max_autotune.py::TestMaxAutotuneRemoteCache::test_get_mm_configs_float32_precision_ieee [GH job link](https://github.com/pytorch/pytorch/actions/runs/17283985553/job/49058214260) [HUD commit link](e9d34b2438) ([comment](https://github.com/pytorch/pytorch/pull/161545#issuecomment-3233569771))
This commit is contained in:
PyTorch MergeBot 2025-08-28 13:46:47 +00:00
parent ef0483d74c
commit 05d0f11dbd
1 changed files with 2 additions and 36 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

38
test/inductor/test_max_autotune.py
View File

@ -28,9 +28,8 @@ from torch._inductor.autotune_process import (
TuningProcessPool, TuningProcessPool,
) )
from torch._inductor.graph import GraphLowering from torch._inductor.graph import GraphLowering
from torch._inductor.ir import Buffer, ChoiceCaller, FixedLayout, InputBuffer from torch._inductor.ir import Buffer, ChoiceCaller, FixedLayout
from torch._inductor.kernel.mm_plus_mm import aten_mm_plus_mm from torch._inductor.kernel.mm_plus_mm import aten_mm_plus_mm
from torch._inductor.kernel_inputs import MMKernelInputs
from torch._inductor.select_algorithm import ( from torch._inductor.select_algorithm import (
add_feedback_saver, add_feedback_saver,
AlgorithmSelectorCache, AlgorithmSelectorCache,
@ -76,7 +75,7 @@ from torch.testing._internal.inductor_utils import (
) )
torch.backends.cuda.matmul.allow_tf32 = True torch.set_float32_matmul_precision("high")
if HAS_CUDA_AND_TRITON: if HAS_CUDA_AND_TRITON:
torch.cuda.memory._set_allocator_settings("expandable_segments:False") torch.cuda.memory._set_allocator_settings("expandable_segments:False")
@ -2077,39 +2076,6 @@ class TestMaxAutotuneRemoteCache(TestCase):
global_stats.report() global_stats.report()
self.assertEqual(global_stats.autotune_remote, Stats(2, 3, 2)) self.assertEqual(global_stats.autotune_remote, Stats(2, 3, 2))
def test_get_mm_configs_float32_precision_ieee(self):
"""Test that configs returned from choices.get_mm_configs use float32_precision == ieee."""
from torch._inductor.choices import InductorChoices
from torch._inductor.graph import GraphLowering
from torch._inductor.ir import FlexibleLayout
from torch.fx.experimental.proxy_tensor import make_fx
# Create a simple graph to get proper context
gm = make_fx(lambda: torch.zeros(2, 3))()
graph = GraphLowering(gm)
with V.set_graph_handler(graph):
device = torch.device(f"{GPU_TYPE}:0")
mat1 = InputBuffer(
name="mat1",
layout=FixedLayout(device, torch.float32, [64, 128], [128, 1]),
)
mat2 = InputBuffer(
name="mat2",
layout=FixedLayout(device, torch.float32, [128, 64], [64, 1]),
)
kernel_inputs = MMKernelInputs([mat1, mat2])
output_layout = FlexibleLayout(device, torch.float32, [64, 64])
choices = InductorChoices()
configs = list(
choices.get_mm_configs(kernel_inputs, output_layout, "mm", "mm")
)
for cfg in configs:
self.assertIn("ALLOW_TF32", cfg)
self.assertEqual(cfg["ALLOW_TF32"], True)
class _TestTritonTemplateCaller(TritonTemplateCaller): class _TestTritonTemplateCaller(TritonTemplateCaller):
def __init__(self, bmreq: _TestBenchmarkRequest): def __init__(self, bmreq: _TestBenchmarkRequest):