[#18362][relax.frontend.torch] change to temporary solution

Author: cchung100m

3rdparty/dlpack

@@ -0,0 +1 @@
+Subproject commit e2bdd3bee8cb6501558042633fa59144cc8b7f5f

3rdparty/vta-hw

@@ -0,0 +1 @@
+Subproject commit 36a91576edf633479c78649e050f18dd2ddc8103

python/tvm/relax/frontend/torch/exported_program_translator.py

@@ -75,10 +75,10 @@ def _randn(self, node: fx.Node) -> relax.Var:
 
         shape = relax.ShapeExpr(size)
 
-        mean = relax.const(0.0, dtype)
-        std = relax.const(1.0, dtype)
-
-        return self.block_builder.emit(relax.op.nn.normal(mean, std, shape, dtype))
+        # TODO: This is a temporary solution that returns zeros instead of random values
+        # since random initialization is mainly used during training, not inference.
+        # This should be updated once Relax adds proper random number generation support.
+        return self.block_builder.emit(relax.op.zeros(shape, dtype))
 
     ########## Neural Network ##########
 

tests/python/relax/test_frontend_from_exported_program.py

@@ -6156,9 +6156,9 @@ def main(
             with R.dataflow():
                 lv0 = R.zeros((5, 5), dtype="float32")
 
-                lv1 = R.random.normal(
-                    R.const(0, "float32"), R.const(1, "float32"), R.shape([5]), dtype="float32"
-                )
+                # Use zeros instead of random normal distribution
+                lv1 = R.zeros((5,), dtype="float32")
+
                 lv2 = R.nn.elu(lv1)
                 lv3 = R.add(lv2, R.const(1.0, "float32"))
                 v = R.add(lv3, R.const(1e-8, "float32"))
@@ -6168,7 +6168,6 @@ def main(
                 )
 
                 L = R.tensor_update(lv0, (idx, idx), v)
-
                 y = R.add(x, R.const(1, "float32"))
 
                 gv = R.tuple(y, L)