Upstream patches\

Golden_Repo/p/PyTorch/PyTorch-1.11-gcccoremkl-11.2.0-2021.4.0-CUDA-11.5.eb

@@ -74,6 +74,10 @@ patches = [
     # 'cub-reduce.patch',
     # 'cub-math-gpu.patch',
     # 'cub-CMake-Dependencies.patch',
+    'PyTorch-1.11.0_fix_skip_jit_cuda_fuser.patch',
+    'PyTorch-1.11.0_increas-distributed-test-timeout.patch',
+    'PyTorch-1.11.0_skip_failing_ops_tests.patch',
+
 
 
 ]
@@ -117,9 +121,6 @@ dependencies = [
     ('expecttest', '0.1.3'),
 ]
 
-# default CUDA compute capabilities to use (override via --cuda-compute-capabilities)
-cuda_compute_capabilities = ['6.0', '6.1', '7.0', '7.2', '7.5', '8.0', '8.6']
-
 custom_opts = ["USE_CUPTI_SO=1"]
 configopts = 'MKL_THREADING_LAYER=sequential CFLAGS="$CFLAGS -fopenmp" CXXFLAGS="$CXXFLAGS -fopenmp" LDFLAGS=-fopenmp'
 
@@ -176,7 +177,6 @@ excluded_tests = {
         'distributed/test_c10d_nccl',  # fails on jureca dc
         'distributed/test_c10d_spawn_nccl',  # fails on jureca dc
         'distributed/test_data_parallel',  # fails on jureca dc
-
     ]
 }
 

Golden_Repo/p/PyTorch/PyTorch-1.11.0_increas-distributed-test-timeout.patch

+It seems the timeout for the distributed tests is set to low and spurious failures can be seen
+Increase it by a factor of 6 similar to torch/testing/_internal/distributed/distributed_test.py
+
+Original patch by Alexander Grund (TU Dresden), updated by Caspar van Leeuwen (SURF)
+
+diff -Nru pytorch-1.11.0-rc3.orig/torch/testing/_internal/common_distributed.py pytorch-1.11.0-rc3/torch/testing/_internal/common_distributed.py
+--- pytorch-1.11.0-rc3.orig/torch/testing/_internal/common_distributed.py	2022-02-24 18:07:16.414274654 +0100
++++ pytorch-1.11.0-rc3/torch/testing/_internal/common_distributed.py	2022-02-24 18:08:31.772851148 +0100
+@@ -321,7 +321,7 @@
+     # TSAN runs much slower.
+     TIMEOUT_DEFAULT = 500
+ else:
+-    TIMEOUT_DEFAULT = 100
++    TIMEOUT_DEFAULT = 600
+ TIMEOUT_OVERRIDE = {"test_ddp_uneven_inputs": 400}
+ 
+ 
\ No newline at end of file

Golden_Repo/p/PyTorch/PyTorch-1.11.0_skip_failing_ops_tests.patch

+diff -Nru pytorch-1.11.0-rc3.orig/torch/testing/_internal/common_methods_invocations.py pytorch-1.11.0-rc3/torch/testing/_internal/common_methods_invocations.py
+--- pytorch-1.11.0-rc3.orig/torch/testing/_internal/common_methods_invocations.py	2022-02-24 18:07:16.430276050 +0100
++++ pytorch-1.11.0-rc3/torch/testing/_internal/common_methods_invocations.py	2022-02-24 19:38:11.610293957 +0100
+@@ -8791,7 +8791,10 @@
+            supports_fwgrad_bwgrad=True,
+            autodiff_fusible_nodes=['aten::contiguous'],
+            assert_jit_shape_analysis=True,
+-           supports_out=False),
++           supports_out=False,
++           skips=(
++               DecorateInfo(unittest.skip("Skipped!"), 'TestJit', 'test_variant_consistency_jit', device_type='cpu'),
++           )),
+     OpInfo('sum_to_size',
+            op=lambda x, *args, **kwargs: x.sum_to_size(*args, **kwargs),
+            dtypes=floating_and_complex_types_and(torch.float16, torch.bfloat16),
+@@ -9746,6 +9749,10 @@
+                DecorateInfo(unittest.skip("Skipped!"), 'TestMathBits', 'test_neg_view', device_type='cuda'),
+                DecorateInfo(unittest.skip("Skipped!"), 'TestCommon', 'test_dtypes'),
+                DecorateInfo(unittest.skip("Skipped!"), 'TestGradients', 'test_fn_gradgrad'),
++               # It also breaks on CPU. We'll revisit this once `linalg.lu_solve` is a thing
++               # See https://github.com/pytorch/pytorch/pull/64387 and https://github.com/pytorch/pytorch/issues/67767
++               DecorateInfo(unittest.skip("Skipped!"), 'TestGradients', 'test_fn_grad',
++                            dtypes=(torch.complex128,)),
+            )),
+     OpInfo('linalg.cholesky',
+            aten_name='linalg_cholesky',
\ No newline at end of file