[FIX] Infer input shape in sparse_dense_padded's alter_op if one does…

… not exist (apache#7308)

* [FIX] Infer input shape in sparse_dense_padded's alter_op if one does not exist

If there are multiple alter_ops in a model, the first alteration does
not run type inference for the subsequent ones. In this case, we don't
have the shape information, so we run the inferencer manually.

* add todo

python/tvm/topi/cuda/sparse.py

@@ -292,7 +292,14 @@ def is_valid_for_sparse_dense_padded(data, weight_data):
     """
     # pylint:disable=invalid-name
     warp_size = int(tvm.target.Target.current(allow_none=False).thread_warp_size)
-    m = get_const_tuple(data.checked_type.shape)[1]
+    # If there are multiple alter_ops in a model, the first alteration does not
+    # run type inference for the subsequent ones. In this case, we don't have
+    # the shape information, so we run the inferencer manually.
+    try:
+        m = get_const_tuple(data.checked_type.shape)[1]
+    except ValueError:
+        data_infered = relay.transform.InferType()(tvm.IRModule.from_expr(data))["main"]
+        m = get_const_tuple(data_infered.ret_type.shape)[1]
     if len(weight_data.shape) == 1:
         bs_m = 1
     else:

src/relay/transforms/alter_op_layout.cc

@@ -110,6 +110,7 @@ class AlterTransformMemorizer : public TransformMemorizer {
  * 2. Do not support nested tuple arguments.
  */
 Expr AlterOpLayout(const Expr& expr) {
+  // TODO(@icemelon9): need to rerun type inference after applying an alter op.
   AlterTransformMemorizer alterMemorizer(make_object<AlterTransformMemorizerNode>());
   auto fcontext = [&](const Call& call) -> ObjectRef { return alterMemorizer; };