accounted for l1_loss now being a composite op

Author: Chillee

functorch/csrc/BatchRulesDecompositions.cpp

@@ -130,6 +130,7 @@ TORCH_LIBRARY_IMPL(aten, FT_BATCHED_KEY, m) {
   OP_DECOMPOSE(inner);
   OP_DECOMPOSE(instance_norm);
   OP_DECOMPOSE(kron);
+  OP_DECOMPOSE(l1_loss);
   OP_DECOMPOSE(layer_norm);
   OP_DECOMPOSE2(ldexp, Tensor);
   OP_DECOMPOSE2(less_equal, Tensor );

test/test_ops.py

@@ -778,7 +778,6 @@ def test_vmapjvpall(self, device, dtype, op):
         xfail('fft.ihfftn'),  # conj_physical fallback
         xfail('istft'),  # col2im fallback
         xfail('polar'),  # complex fallback
-        xfail('nn.functional.l1_loss', ''),
         xfail('nn.functional.max_unpool3d', 'grad'),
         xfail('nn.functional.smooth_l1_loss', ''),
         xfail('nn.functional.max_unpool2d', 'grad'),
@@ -911,7 +910,6 @@ def test():
         xfail('nn.functional.soft_margin_loss', ''),
         xfail('scatter_reduce', 'amin'),
         xfail('nn.functional.max_unpool1d', 'grad'),
-        xfail('nn.functional.l1_loss', ''),
         xfail('nn.functional.max_unpool2d', 'grad'),
         xfail('qr'),
         xfail('linalg.eigvalsh'),  # _linalg_eigh doesn't have batching rule

test/test_vmap.py

@@ -3267,7 +3267,6 @@ def test_vmap_exhaustive(self, device, dtype, op):
         xfail('nn.functional.triplet_margin_loss', ''),
         xfail('nn.functional.pdist', ''),
         xfail('scatter_reduce', 'sum'),
-        xfail('nn.functional.l1_loss', ''),
         xfail('nn.functional.smooth_l1_loss', ''),
         xfail('scatter_reduce', 'amax'),
         xfail('nn.functional.max_unpool1d', 'grad'),