Remove some xfails for transform coverage (#910)

Author: zou3519

functorch/csrc/BatchRulesLinearAlgebra.cpp

@@ -197,7 +197,7 @@ TORCH_LIBRARY_IMPL(aten, FT_BATCHED_KEY, m) {
   VARIADIC_BDIMS_BOXED(symeig);
   VARIADIC_BDIMS_BOXED(triangular_solve);
 
-  VARIADIC_BDIMS_BOXED(_det_lu_based_helper);
+  VARIADIC_BDIMS_BOXED(_linalg_det);
   VARIADIC_BDIMS_BOXED(_lu_with_info);
 }
 }}

functorch/csrc/PyTorchOperatorHacks.cpp

@@ -212,7 +212,7 @@ Tensor binary_cross_entropy_with_logits_hack(
   const Tensor& pos_weight = c10::value_or_else(pos_weight_opt, [] {return Tensor();});
 
   Tensor loss;
-  auto max_val = (-input).clamp_min_(0);
+  auto max_val = (-input).clamp_min(0);
   if (pos_weight.defined()) {
     // pos_weight need to be broadcasted, thus mul(target) is not inplace.
     auto log_weight = (pos_weight - 1).mul(target).add_(1);

test/test_ops.py

@@ -379,11 +379,6 @@ def wrapped_fn(*args, **kwargs):
         skip('svd_lowrank', ''),  # fails on cuda, runs okay on cpu
         skip('nn.functional.dropout2d', ''),  # fails on cuda, runs okay on cpu
 
-        # The following don't have a forward-mode AD formula in PyTorch core
-        # (check derivatives.yaml).
-        xfail('var_mean'),
-        xfail('std_mean'),
-
         # =============================================
         # NB: The above failures also fail using PyTorch core's
         #     forward-mode AD and vmap.
@@ -674,14 +669,11 @@ def test_vmapvjp(self, device, dtype, op):
         # skip because this is flaky depending on what the max_norm is!
         skip('nn.functional.embedding', ''),
         xfail('nn.functional.soft_margin_loss', ''),
-        xfail('nn.functional.binary_cross_entropy_with_logits', ''),
         xfail('linalg.householder_product'),
         xfail('tensor_split'),
         xfail('quantile'),
-        xfail('var_mean'),
         xfail('as_strided'),
         xfail('nn.functional.gaussian_nll_loss'),
-        xfail('std_mean'),
         xfail('scatter'),
         xfail('matrix_exp'),
         xfail('nanquantile'),
@@ -765,6 +757,8 @@ def test_vmapjvpall(self, device, dtype, op):
         xfail('nn.functional.max_pool3d'),
         xfail('vdot'),
         xfail('linalg.cross'),
+        xfail('nanmean'),
+        xfail('nansum'),
         xfail('nn.functional.feature_alpha_dropout', 'without_train'),
         xfail('linalg.lu_factor', ''),
         xfail('nn.functional.dropout2d', ''),
@@ -782,7 +776,6 @@ def test_vmapjvpall(self, device, dtype, op):
         xfail('nn.functional.smooth_l1_loss', ''),
         xfail('nn.functional.max_unpool2d', 'grad'),
         xfail('nn.functional.soft_margin_loss', ''),
-        xfail('nn.functional.binary_cross_entropy_with_logits', ''),
         xfail('nn.functional.max_unpool1d', 'grad'),
         xfail('nn.functional.embedding', ''),
         xfail('lu_unpack'),
@@ -1044,23 +1037,16 @@ def get_vjp(cotangents, *primals):
         # RuntimeError: Trying to set a forward gradient that has a different size than that of the original Tensor,
         # this is not supported. Tensor is of size [5, 2, 3] while the given forward gradient is of size [1, 2, 3].
         xfail('normal', ''),
-        xfail('_masked.amax', ''),
-        xfail('_masked.amin', ''),
         xfail('_masked.log_softmax', ''),
         xfail('_masked.softmax', ''),
         xfail('_masked.softmin', ''),
-        xfail('amax', ''),
-        xfail('amin', ''),
         xfail('cdist', ''),
         xfail('cholesky', ''),
         xfail('eig', ''),
         xfail('linalg.det', ''),
-        xfail('linalg.matrix_norm', ''),
         xfail('linalg.slogdet', ''),
         xfail('logcumsumexp', ''),
         xfail('logdet', ''),
-        xfail('nanmean', ''),
-        xfail('nansum', ''),
         xfail('nn.functional.embedding_bag', ''),
         xfail('nn.functional.grid_sample', ''),
         xfail('nn.functional.hardsigmoid', ''),
@@ -1070,9 +1056,7 @@ def get_vjp(cotangents, *primals):
         xfail('nn.functional.softmin', ''),
         xfail('nn.functional.softmin', 'with_dtype'),
         xfail('renorm', ''),
-        xfail('std_mean', ''),
         xfail('symeig', ''),
-        xfail('var_mean', ''),
         xfail('nn.functional.feature_alpha_dropout', 'with_train'),
         xfail('nn.functional.kl_div', ''),
         xfail('pca_lowrank', ''),
@@ -1090,7 +1074,6 @@ def get_vjp(cotangents, *primals):
         xfail('scatter_reduce', 'mean'),
         xfail('scatter_reduce', 'prod'),
         skip('linalg.householder_product', '', device_type='cuda'),  # flaky, I'm not sure why
-        xfail('nn.functional.binary_cross_entropy_with_logits'),
     }))
     def test_jvpvjp(self, device, dtype, op):
         if not op.supports_autograd:

test/test_vmap.py

@@ -3177,6 +3177,7 @@ def test_vmap_exhaustive(self, device, dtype, op):
         xfail('histogram'),
         xfail('index_fill'),
         xfail('nansum'),
+        xfail('nanmean'),
         # `index_put` OpInfo in pytorch/pytorch has
         # masked index as input which is not supported
         xfail('index_put', ''),