From 6c4412f72b6bb017e24623d4c462abb29b7a47b6 Mon Sep 17 00:00:00 2001
From: PyTorch MergeBot <pytorchmergebot@users.noreply.github.com>
Date: Tue, 21 Oct 2025 17:44:42 +0000
Subject: [PATCH] Revert "[Inductor] support masked vectorization for the
 tail_loop for float64 datatype (#163316)"

This reverts commit e9d89734274a4a2640fa77b898c800a87d1d874e.

Reverted https://github.com/pytorch/pytorch/pull/163316 on behalf of https://github.com/clee2000 due to seems to have broken some no_gpu tests? test/inductor/test_cpu_repro.py::CPUReproTests::test_double_reduction_vec [GH job link](https://github.com/pytorch/pytorch/actions/runs/18689033019/job/53290772740) [HUD commit link](https://hud.pytorch.org/pytorch/pytorch/commit/e9d89734274a4a2640fa77b898c800a87d1d874e) ([comment](https://github.com/pytorch/pytorch/pull/163316#issuecomment-3428210509))
---
 test/inductor/test_cpu_repro.py | 64 ---------------------------------
 torch/_inductor/codegen/cpp.py  |  1 -
 2 files changed, 65 deletions(-)

diff --git a/test/inductor/test_cpu_repro.py b/test/inductor/test_cpu_repro.py
index e887367dd8d..32e8049dd7f 100644
--- a/test/inductor/test_cpu_repro.py
+++ b/test/inductor/test_cpu_repro.py
@@ -4810,22 +4810,6 @@ class CPUReproTests(TestCase):
         self.common(fn, (x,))
         check_metrics_vec_kernel_count(1)
 
-        # Tail vectorization case
-        x = torch.randn((22, 22), dtype=torch.double)
-        torch._dynamo.reset()
-        metrics.reset()
-        with torch.no_grad():
-            expected = fn(x)
-            compiled_fn = torch.compile(fn)
-            actual, code = run_and_get_cpp_code(compiled_fn, x)
-            self.assertEqual(expected, actual)
-            # 1 generated vec kernel
-            self.assertEqual(metrics.generated_cpp_vec_kernel_count, 1)
-            # Check that both main and tail loops are vectorized
-            FileCheck().check_count(
-                "at::vec::VectorizedN<double,2>::loadu", 2, exactly=True
-            ).run(code)
-
     def test_double_reduction_vec(self):
         def fn(x):
             return x.sum(dim=1)
@@ -4835,22 +4819,6 @@ class CPUReproTests(TestCase):
         self.common(fn, (x,))
         check_metrics_vec_kernel_count(1)
 
-        # Tail vectorization case
-        x = torch.randn((22, 22), dtype=torch.double)
-        torch._dynamo.reset()
-        metrics.reset()
-        with torch.no_grad():
-            expected = fn(x)
-            compiled_fn = torch.compile(fn)
-            actual, code = run_and_get_cpp_code(compiled_fn, x)
-            self.assertEqual(expected, actual)
-            # 1 generated vec kernel
-            self.assertEqual(metrics.generated_cpp_vec_kernel_count, 1)
-            # Check that both main and tail loops are vectorized
-            FileCheck().check_count(
-                "at::vec::VectorizedN<double,2>::loadu", 2, exactly=True
-            ).run(code)
-
     def test_convert_fp32_to_double_vec(self):
         def fn(x):
             return x.to(torch.double)
@@ -4860,22 +4828,6 @@ class CPUReproTests(TestCase):
         self.common(fn, (x,))
         check_metrics_vec_kernel_count(1)
 
-        # Tail vectorization case
-        x = torch.randn(22, 22)
-        torch._dynamo.reset()
-        metrics.reset()
-        with torch.no_grad():
-            expected = fn(x)
-            compiled_fn = torch.compile(fn)
-            actual, code = run_and_get_cpp_code(compiled_fn, x)
-            self.assertEqual(expected, actual)
-            # 1 generated vec kernel
-            self.assertEqual(metrics.generated_cpp_vec_kernel_count, 1)
-            # Check that both main and tail loops are vectorized
-            FileCheck().check_count(
-                "at::vec::convert<double,2,float,1>", 2, exactly=True
-            ).run(code)
-
     def test_convert_double_to_fp32_vec(self):
         def fn(x):
             return x.to(torch.float32)
@@ -4885,22 +4837,6 @@ class CPUReproTests(TestCase):
         self.common(fn, (x,))
         check_metrics_vec_kernel_count(1)
 
-        # Tail vectorization case
-        x = torch.randn((22, 22), dtype=torch.double)
-        torch._dynamo.reset()
-        metrics.reset()
-        with torch.no_grad():
-            expected = fn(x)
-            compiled_fn = torch.compile(fn)
-            actual, code = run_and_get_cpp_code(compiled_fn, x)
-            self.assertEqual(expected, actual)
-            # 1 generated vec kernel
-            self.assertEqual(metrics.generated_cpp_vec_kernel_count, 1)
-            # Check that both main and tail loops are vectorized
-            FileCheck().check_count(
-                "at::vec::convert<float,1,double,2>", 2, exactly=True
-            ).run(code)
-
     def test_no_redundant_to_dtypes_between_fused_scheduler_node(self):
         # https://github.com/pytorch/pytorch/issues/115260
         p0 = torch.tensor([1.0879], dtype=torch.float16)
diff --git a/torch/_inductor/codegen/cpp.py b/torch/_inductor/codegen/cpp.py
index 6602a821c56..1b8b0a9b9e2 100644
--- a/torch/_inductor/codegen/cpp.py
+++ b/torch/_inductor/codegen/cpp.py
@@ -159,7 +159,6 @@ VECTORIZABLE_DTYPES: list[torch.dtype] = [
 ]
 
 MASKED_VECTORIZABLE_DTYPES: list[torch.dtype] = [
-    torch.float64,
     torch.float,
     torch.bfloat16,
     torch.float16,