diff --git a/python/tvm/relax/frontend/torch/base_fx_graph_translator.py b/python/tvm/relax/frontend/torch/base_fx_graph_translator.py
index 3ff3b596af3b..6f5ffdf4aba7 100644
--- a/python/tvm/relax/frontend/torch/base_fx_graph_translator.py
+++ b/python/tvm/relax/frontend/torch/base_fx_graph_translator.py
@@ -2120,29 +2120,31 @@ def _chunk(self, node: fx.Node) -> relax.Var:
             relax.op.split(x=x, indices_or_sections=n_sections, axis=dim)
         )
 
-    def _cumprod(self, node: fx.Node) -> relax.Var:
-        x = self.env[node.args[0]]
+    def _cumulative_dtype(self, x: relax.Expr, node: fx.Node) -> str | None:
+        """The output dtype of torch.cumsum / torch.cumprod for ``x``.
 
-        dim = node.args[1] if len(node.args) > 1 else node.kwargs.get("dim", None)
+        With no ``dtype`` argument torch accumulates every integral and bool input in
+        int64 -- ``uint8 [200, 100].cumsum(0)`` is int64 ``[200, 300]``, not a uint8 that
+        wraps to ``[200, 44]`` -- and keeps floating inputs as they are.
+        """
         if "dtype" in node.kwargs:
-            dtype = self._convert_data_type(str(node.kwargs["dtype"]), self.env)
-        else:
-            dtype = None
+            return self._convert_data_type(str(node.kwargs["dtype"]), self.env)
+        dtype = x.ty.dtype
+        if dtype.matches_code(DataTypeCode.INT, DataTypeCode.UINT) or str(dtype) == "bool":
+            return "int64"
+        return None
 
-        return self.block_builder.emit(relax.op.cumprod(x, dim, dtype))
+    def _cumprod(self, node: fx.Node) -> relax.Var:
+        x = self.env[node.args[0]]
+        dim = node.args[1] if len(node.args) > 1 else node.kwargs.get("dim", None)
+        return self.block_builder.emit(relax.op.cumprod(x, dim, self._cumulative_dtype(x, node)))
 
     def _cumsum(self, node: fx.Node) -> relax.Var:
         x = self.env[node.args[0]]
-
         dim = node.args[1] if len(node.args) > 1 else node.kwargs.get("dim", None)
-        if "dtype" in node.kwargs:
-            dtype = self._convert_data_type(str(node.kwargs["dtype"]), self.env)
-        else:
-            dtype = None
         if "out" in node.kwargs:
             raise ValueError("specifying out for cumsum is not supported yet")
-
-        return self.block_builder.emit(relax.op.cumsum(x, dim, dtype))
+        return self.block_builder.emit(relax.op.cumsum(x, dim, self._cumulative_dtype(x, node)))
 
     def _expand(self, node: fx.Node) -> relax.Var:
         args = self.retrieve_args(node)
diff --git a/tests/python/relax/test_frontend_from_exported_program.py b/tests/python/relax/test_frontend_from_exported_program.py
index d78629f5737b..fcfc41cca50f 100644
--- a/tests/python/relax/test_frontend_from_exported_program.py
+++ b/tests/python/relax/test_frontend_from_exported_program.py
@@ -5037,6 +5037,59 @@ def main(input_1: R.Tensor((1, 2, 3, 4), dtype="float32")) -> R.Tuple(
     verify_model(Cumsum(), example_args, {}, expected1)
 
 
+def test_cumsum_integer_input_accumulates_in_int64():
+    # With no dtype argument torch accumulates every integral and bool input in int64.
+    # Keeping the input dtype instead makes the running sum wrap: uint8 [200, 100, 50]
+    # is [200, 44, 94] rather than [200, 300, 350].
+    class Cumsum(Module):
+        def forward(self, x):
+            return torch.cumsum(x, dim=1)
+
+    @tvm.script.ir_module
+    class expected:
+        @R.function
+        def main(x: R.Tensor((2, 3), dtype="uint8")) -> R.Tuple(R.Tensor((2, 3), dtype="int64")):
+            with R.dataflow():
+                lv: R.Tensor((2, 3), dtype="int64") = R.cumsum(x, axis=1, dtype="int64")
+                gv: R.Tuple(R.Tensor((2, 3), dtype="int64")) = (lv,)
+                R.output(gv)
+            return gv
+
+    x = torch.tensor([[200, 100, 50], [2, 3, 4]], dtype=torch.uint8)
+    verify_model(Cumsum(), (x,), {}, expected)
+
+
+@pytest.mark.parametrize("dtype", [torch.bool, torch.uint8, torch.int8, torch.int32, torch.int64])
+def test_cumsum_cumprod_integer_values(dtype):
+    class Cumsum(Module):
+        def forward(self, x):
+            return torch.cumsum(x, dim=1)
+
+    class Cumprod(Module):
+        def forward(self, x):
+            return torch.cumprod(x, dim=1)
+
+    class CumsumFloat(Module):
+        def forward(self, x):
+            return torch.cumsum(x, dim=0, dtype=torch.float32)
+
+    if dtype is torch.bool:
+        x = torch.tensor([[True, True, False], [True, True, True]])
+    elif dtype is torch.uint8:
+        x = torch.tensor([[200, 100, 50], [2, 3, 4]], dtype=dtype)
+    elif dtype is torch.int8:
+        x = torch.tensor([[100, 100, 50], [2, 3, 4]], dtype=dtype)
+    else:
+        x = torch.tensor([[2**30, 2**30, 5], [2, 3, 4]], dtype=dtype)
+    for model in (Cumsum(), Cumprod(), CumsumFloat()):
+        with torch.no_grad():
+            want = model(x)
+        mod = from_exported_program(export(model, (x,)))
+        got_dtype = str(mod["main"].ret_ty.fields[0].dtype)
+        assert got_dtype == str(want.dtype).replace("torch.", "")
+        verify_model_numerically(model, (x,))
+
+
 def test_expand():
     class Expand1(Module):
         def forward(self, x):