Merge #618

618: fix litmus bug when using muliple inputs r=dkim-furiosa a=maxstate

### Problem
* litmus를 할시 input의 갯수를 하나로 가정하고 있었습니다.

### Solution
* 여러 input을 사용할 수 있도록 합니다.

### Testing
* 기존 테스트 통과

### Checklist
- [x] Will this be part of a product update? If yes, please update [Changelog](https://github.com/furiosa-ai/furiosa-sdk-private/blob/main/CHANGELOG.md).
- [x] Changes are less than 1000 lines (refer to [Pull Request 크기 줄이는 법](https://github.com/furiosa-ai/npu-tools/blob/master/CodingConvention.md#pull-request-%ED%81%AC%EA%B8%B0-%EC%A4%84%EC%9D%B4%EB%8A%94-%EB%B2%95))
- [x] My code has passed black/isort.
- [x] I have performed a self-review of my code.
- [x] All CI checks have passed.


Co-authored-by: SH. Song <[email protected]>

python/furiosa-litmus/furiosa/litmus/__init__.py

@@ -33,6 +33,7 @@ def calibrate_with_random_data(
     initializers = set(tensor.name for tensor in model.graph.initializer)
     rng = np.random.default_rng()
     for _ in range(dataset_size):
+        inputs = []
         for value_info in model.graph.input:
             if value_info.name in initializers:
                 continue
@@ -77,18 +78,18 @@ def calibrate_with_random_data(
                     )
             np_dtype = onnx.mapping.TENSOR_TYPE_TO_NP_TYPE[value_info.type.tensor_type.elem_type]
             if np.issubdtype(np_dtype, np.floating):
-                inputs = rng.standard_normal(size=shape, dtype=np_dtype)
+                inputs.append(rng.standard_normal(size=shape, dtype=np_dtype))
             elif np.issubdtype(np_dtype, np.integer):
                 iinfo = np.iinfo(np_dtype)
-                inputs = rng.integers(
-                    iinfo.min, iinfo.max, size=shape, dtype=np_dtype, endpoint=True
+                inputs.append(
+                    rng.integers(iinfo.min, iinfo.max, size=shape, dtype=np_dtype, endpoint=True)
                 )
             else:
                 elem_type = onnx.TensorProto.DataType.Name(value_info.type.tensor_type.elem_type)
                 raise NotImplementedError(
                     f"tensor '{value_info.name}' is of {elem_type} but a model whose input tensor is of {elem_type} cannot be randomly calibrated yet"
                 )
-        calibrator.collect_data([[inputs]])
+        calibrator.collect_data([inputs])
     return calibrator.compute_range()