Small. Fix test in DebugMode (the transfer was complaining about inf values,…

theano/gpuarray/blas.py

@@ -1735,6 +1735,7 @@ def local_inplace_gpuagemm(node, inputs):
 def local_inplace_gpuager(node, inputs):
     return [gpuger_inplace(*inputs)]
 
+
 @inplace_allocempty(GpuGemmBatch, 0)
 def local_inplace_gpuagemmbatch(node, inputs):
     return [gpugemmbatch_inplace(*inputs)]

theano/gpuarray/opt.py

@@ -1235,7 +1235,8 @@ def local_gpua_gemmbatch(op, context_name, inputs, outputs):
         if b.dtype != out_dtype:
             b = gpu_cast_op(b)
 
-    c = tensor.AllocEmpty(out_dtype)(a.shape[0], a.shape[1], b.shape[2])
+    c = GpuAllocEmpty(out_dtype, context_name)(
+        a.shape[0], a.shape[1], b.shape[2])
     out = gpugemmbatch_no_inplace(c, np.asarray(1.0, dtype=out_dtype),
                                   a, b, np.asarray(0.0, dtype=out_dtype))
     if len(output_dims) != 3:

theano/gpuarray/tests/test_blas.py

@@ -166,6 +166,7 @@ class TestGpuGemmBatchStrided(TestCase):
         x_num = np.arange(32 * 19 * 600, dtype=config.floatX).reshape((32, 19, 600))
         y_num = np.arange(7 * 32 * 600, dtype=config.floatX).reshape((32, 7, 600))
         f(x_num, y_num)
+        assert f.maker.fgraph.toposort()[-2].op.inplace
 
 
 class TestGpuSger(TestGer):