[release/2.6] [SWDEV-529824] Fix Float16 CooperativeReduction Test Failure (#2204)

- Previously expected values were calculated on GPU using same dtype as
result values
- Now expected values are calculated on CPU using Float32 dtype
- This fixes a test failure that was observed on Navi48 where difference
between Eager mode (expected) and Inductor / Triton (result) did not
meet the error tolerance when sum was evaluated on an array of Float16
values

Co-authored-by: pnikolic-amd <[email protected]>
(cherry picked from commit 8fe3cdd)
(cherry picked from commit 34f3b3e)

Author: 2 people

test/inductor/test_cooperative_reductions.py

@@ -58,7 +58,8 @@ def setUp(self):
         torch._dynamo.reset()
 
     def run_and_check(self, fn, args, *, expect_kernel_count=1):
-        expected = fn(*args)
+        args_cpu = [tensor.cpu().to(torch.float32) for tensor in args]
+        expected = fn(*args_cpu).to(torch.float16)
         fn = torch.compile(fn, fullgraph=True)
         result, (source_code,) = run_and_get_code(fn, *args)
         self.assertEqual(result, expected)