ALinrunrun opened a new issue, #19564:
URL: https://github.com/apache/tvm/issues/19564
### Expected behavior
TVM Relax should execute ONNX `ScatterElements` consistently with ONNX
Runtime when `reduction="add"` is used.
For duplicated target indices, the updates should be accumulated with the
original input value.
### Actual behavior
TVM appears to overwrite with the last update value instead of applying add
reduction:
```
onnxruntime:
[[16. 2. 2.]
[ 2. 2. 2.]
[ 2. 20. 2.]
[ 2. 2. 26.]]
tvm:
[[11. 2. 2.]
[ 2. 2. 2.]
[ 2. 13. 2.]
[ 2. 2. 17.]]
max_abs_diff: 9.0
```
### Environment
TVM: 0.14 environment / Relax ONNX frontend
ONNX Runtime: 1.23
Python: 3.11
Target: llvm
OS: Linux
### Steps to reproduce
```
import numpy as np
import onnx
import onnxruntime as ort
import tvm
from onnx import TensorProto, helper
from tvm import relax
from tvm.relax.frontend.onnx import from_onnx
data = np.full((4, 3), 2.0, dtype=np.float32)
indices = np.array([[0, 2, 3], [0, 2, 3]], dtype=np.int64)
updates = np.array([[3.0, 5.0, 7.0], [11.0, 13.0, 17.0]], dtype=np.float32)
x = helper.make_tensor_value_info("X", TensorProto.FLOAT, [4, 3])
i = helper.make_tensor_value_info("I", TensorProto.INT64, [2, 3])
u = helper.make_tensor_value_info("U", TensorProto.FLOAT, [2, 3])
y = helper.make_tensor_value_info("Y", TensorProto.FLOAT, [4, 3])
node = helper.make_node(
"ScatterElements",
["X", "I", "U"],
["Y"],
axis=0,
reduction="add",
)
graph = helper.make_graph([node], "g", [x, i, u], [y])
model = helper.make_model(graph, opset_imports=[helper.make_opsetid("", 18)])
model.ir_version = 8
model = onnx.shape_inference.infer_shapes(model)
ort_out = ort.InferenceSession(
model.SerializeToString(),
providers=["CPUExecutionProvider"],
).run(None, {"X": data, "I": indices, "U": updates})[0]
mod = from_onnx(model)
mod = relax.transform.LegalizeOps()(mod)
exe = tvm.relax.build(mod, target="llvm")
vm = tvm.relax.VirtualMachine(exe, device=tvm.cpu())
tvm_out = vm["main"](
tvm.runtime.tensor(data, device=tvm.cpu()),
tvm.runtime.tensor(indices, device=tvm.cpu()),
tvm.runtime.tensor(updates, device=tvm.cpu()),
).numpy()
print("onnxruntime:")
print(ort_out)
print("tvm:")
print(tvm_out)
print("max_abs_diff:", float(np.max(np.abs(ort_out - tvm_out))))
```
### Triage
* needs-triage
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
