[FRONTEND] Expose end-to-end compile to python frontend (#58)

python/tests/test_compiler.py

@@ -0,0 +1,23 @@
+import torch
+
+import triton
+import triton.language as tl
+
+# trigger the torch.device implicitly to ensure cuda context initialization
+torch.zeros([10], device=torch.device('cuda'))
+
+
+def test_empty_kernel_cubin_compile():
+    @triton.jit
+    def kernel(X, stride_xm, stride_xn, BLOCK: tl.constexpr):
+        pass
+
+    device = torch.cuda.current_device()
+    cubin = triton.compile(kernel,
+                           "*fp32,i32,i32",
+                           device=device,
+                           constants={"BLOCK": 256},
+                           output="cubin")
+
+    print('cubin size:', len(cubin))
+    assert len(cubin) > 0