[Relay] Remove DynamicToStatic pass from graph runtime build (#10691)

Closes #10692

To solve this problem, we can either remove this pass from `relay.build(...)` pipeline or run `DynamicToStatic` in both VM and non-VM paths. I propose to remove it because  (1) usually `DynamicToStatic` is supposed to be applied after model import and (2) the only case running `DynamicToStatic` during `relay.build(...)` helps is when the input is entirely static but a frontend fails to produce a static mod AND a user forgets to run `DynamicToStatic` after model import.

 I hope the latter case happens rarely but if not, that's something we should fix in the frontend side. We should avoid relying on `DynamicToStatic` that runs during `relay.build(...)` since not all use cases of TVM use `relay.build(...)` (BYOC, for example).

Author: masahi

python/tvm/relay/frontend/paddlepaddle.py

@@ -1281,7 +1281,7 @@ def convert_prelu(g, op, block):
             shape = _op.strided_slice(shape_of(x), [0], [1])
         else:
             shape = _op.strided_slice(shape_of(x), [1], [2])
-        alpha = _op.broadcast_to(alpha, shape)
+        alpha = _op.broadcast_to(alpha, fold_constant(shape))
     out = _op.nn.prelu(x, alpha, axis)
     g.add_node(op.output("Out")[0], out)
 

python/tvm/relay/frontend/pytorch.py

@@ -641,7 +641,7 @@ def full_impl(self, data, fill_value, dtype):
                 tmp.append(_op.cast(_op.expand_dims(dim, axis=0), "int64"))
             size = _op.concatenate(tmp, axis=0)
 
-        out = _op.full(_expr.const(fill_value), size, dtype=dtype)
+        out = _op.full(_expr.const(fill_value, dtype=dtype), size, dtype=dtype)
         if need_reshape:
             out = _op.reshape(out, new_shape)
         return out

python/tvm/relay/frontend/tflite.py

@@ -33,7 +33,7 @@
 from ..backend.name_transforms import sanitize_name
 from .common import ExprTable
 from .common import infer_shape as _infer_shape
-from .common import to_int_list
+from .common import to_int_list, shape_of
 from .tflite_flexbuffer import FlexBufferDecoder
 
 __all__ = ["from_tflite"]
@@ -846,7 +846,7 @@ def convert_shape(self, op):
         input_tensors = self.get_input_tensors(op)
         assert len(input_tensors) == 1, "input tensors length should be 1"
 
-        out = _op.shape_of(self.get_tensor_expr(input_tensors[0]))
+        out = shape_of(self.get_tensor_expr(input_tensors[0]))
 
         return out
 

src/relay/backend/utils.cc

@@ -226,16 +226,6 @@ Array<Pass> GetPassPrefix(bool is_homegeneous, bool is_vm) {
     // eta expand to support constructors in argument position
     pass_seqs.push_back(transform::EtaExpand(
         /* expand_constructor */ true, /* expand_global_var */ false));
-  } else {
-    // DynamicToStatic runs FoldConstant, which affects SimplifyExpr below.
-    // Task extraction uses the is_vm=true branch, meaning SimplifyExpr sees different
-    // inputs from the ones when invoked via relay.build(...).
-    // This causes workload lookups in ApplyHistoryBest to fail if the lookup depends on
-    // the structual hash of the input relay module (e.g. MetaScheduler).
-    // TODO(masahi): Either remove DynamicToStatic below or always run it
-
-    // Convert Dynamic ops to static versions
-    pass_seqs.push_back(transform::DynamicToStatic());
   }
 
   PackedFunc fskip = PackedFunc([](TVMArgs args, TVMRetValue* rv) {
@@ -252,12 +242,12 @@ Array<Pass> GetPassPrefix(bool is_homegeneous, bool is_vm) {
     *rv = false;
   });
   pass_seqs.push_back(transform::EliminateCommonSubexpr(fskip));
-  pass_seqs.push_back(transform::SimplifyExpr());
   pass_seqs.push_back(transform::CombineParallelConv2D(3));
   pass_seqs.push_back(transform::CombineParallelDense(3));
   pass_seqs.push_back(transform::CombineParallelBatchMatmul(3));
   pass_seqs.push_back(transform::FoldConstant());
   pass_seqs.push_back(transform::FoldScaleAxis());
+  pass_seqs.push_back(transform::SimplifyExpr());
   pass_seqs.push_back(transform::CanonicalizeCast());
   pass_seqs.push_back(transform::CanonicalizeOps());
 

tests/python/frontend/tensorflow/test_forward.py

@@ -147,6 +147,7 @@ def run_tvm_graph(
         outputs=out_names,
         convert_config=convert_config,
     )
+
     dev = tvm.device(target, 0)
     if mode == "debug":
         inputs = []
@@ -2421,10 +2422,11 @@ def _test_sparse_to_dense(sparse_indices, sparse_values, default_value, output_s
         )
         oshape = tf.constant(output_shape, shape=output_shape.shape, dtype=str(output_shape.dtype))
 
+        # Output shape depends on a dynamic input, use VM.
         if default_value == None:
             output = tf.sparse_to_dense(indices, oshape, values)
             compare_tf_with_tvm(
-                [sparse_indices, sparse_values], ["indices:0", "values:0"], output.name
+                [sparse_indices, sparse_values], ["indices:0", "values:0"], output.name, mode="vm"
             )
         else:
             dv = tf.placeholder(shape=(), dtype=str(default_value.dtype), name="default_value")
@@ -2433,6 +2435,7 @@ def _test_sparse_to_dense(sparse_indices, sparse_values, default_value, output_s
                 [sparse_indices, sparse_values, default_value],
                 ["indices:0", "values:0", "default_value:0"],
                 output.name,
+                mode="vm",
             )
 
 
@@ -2494,7 +2497,8 @@ def _test_sparse_to_dense_v2(indices, values, A_shape, dtype, default_value=None
 
         result = tf.sparse.to_dense(A_sp, default_value=default_value)
 
-        compare_tf_with_tvm([], [], result.name)
+        # The output shape depends on a dynamic input, use VM.
+        compare_tf_with_tvm([], [], result.name, mode="vm")
 
 
 def test_forward_sparse_to_dense_v2():
@@ -5572,7 +5576,7 @@ def _test_unique(n, dtype, is_dyn):
         if is_dyn:
             compare_tf_with_tvm(np_data, "in_data:0", ["Unique:0", "Unique:1"], mode="vm")
         else:
-            compare_tf_with_tvm(None, "", ["Unique:0", "Unique:1"])
+            compare_tf_with_tvm(np_data, "", ["Unique:0", "Unique:1"], mode="vm")
 
 
 def test_forward_unique():
@@ -5607,7 +5611,10 @@ def _test_unique_with_counts(n, dtype, is_dyn):
             )
         else:
             compare_tf_with_tvm(
-                None, "", ["UniqueWithCounts:0", "UniqueWithCounts:1", "UniqueWithCounts:2"]
+                np_data,
+                "",
+                ["UniqueWithCounts:0", "UniqueWithCounts:1", "UniqueWithCounts:2"],
+                mode="vm",
             )