feat: Add ONNX data to QNN profiler metrics

[xieofxie]

Summary

• Add env-gated ONNX metadata enrichment for QNN basic profiler metrics via WINMLCLI_OP_ADD_DATA
• Include ONNX node type, attributes, and input tensor metadata keyed by input name
• Add focused unit coverage for default and env-enabled output shapes

Verification

• uv run ruff check --fix src\winml\modelkit\optracing\result.py src\winml\modelkit\optracing\qnn\profiler.py tests\unit\optracing\test_qnn_profiler.py
• uv run pytest tests\unit\optracing\test_qnn_profiler.py

[xieofxie]

Sample result

    {
      "name": "/resnet/encoder/stages.1/layers.0/layer/layer.1/convolution/Conv",
      "op_path": "/resnet/encoder/stages.1/layers.0/layer/layer.1/convolution/Conv",
      "op_id": 588,
      "duration_us": 45.286626051604955,
      "percent_of_total": 4.619123992288399,
      "onnx_op_type": "Conv",
      "onnx_attributes": {
        "dilations": [
          1,
          1
        ],
        "group": 1,
        "kernel_shape": [
          3,
          3
        ],
        "pads": [
          1,
          1,
          1,
          1
        ],
        "strides": [
          2,
          2
        ]
      },
      "onnx_inputs": {
        "X": {
          "name": "/resnet/encoder/stages.1/layers.0/layer/layer.0/activation/Relu_output_0_DequantizeLinear_Output",
          "data_type": "FLOAT",
          "dims": [
            1,
            128,
            56,
            56
          ]
        },
        "W": {
          "name": "onnx::Conv_533_DequantizeLinear_Output",
          "data_type": "FLOAT",
          "dims": [
            128,
            128,
            3,
            3
          ]
        },
        "B": {
          "name": "onnx::Conv_534",
          "data_type": "FLOAT",
          "dims": [
            128
          ]
        }
      }
    },

[DingmaomaoBJTU]

Reviewed the ONNX metadata enrichment for QNN basic profiler metrics. The schema-name input resolution, the env gating, and the on/off test coverage look solid. Left a few inline notes — the main one is that an enrichment error can currently void an otherwise-successful trace.

[DingmaomaoBJTU]

When WINMLCLI_OP_ADD_DATA is set, any exception raised inside _load_onnx_operator_data propagates out of _from_csv and discards profiling results that were already computed successfully. Several paths here can throw: load_onnx (missing/corrupt model, or a *_ctx.onnx context binary), TensorProto.DataType.Name() on an unknown enum value, and _serialize_attribute on unsupported attribute types.

Since this is opt-in enrichment, a metadata failure shouldn't be able to void an otherwise-valid trace. Suggest wrapping the call in try/except, logging a warning, and falling back to None.

Verified locally: with the env var set and an unreadable model path, _from_csv raises FileNotFoundError instead of returning the (valid) operator metrics.

[DingmaomaoBJTU]

This fallthrough returns the raw protobuf value for any attribute type not handled above — notably TYPE_PROTO / TYPE_PROTOS (and UNDEFINED). Those are not JSON-serializable, so OpTraceResult.to_json() fails with TypeError: Object of type TypeProto is not JSON serializable.

Verified locally with a TYPE_PROTO attribute. Suggest handling those cases explicitly, or using a safe fallback such as str(value), so serialization can't break.

[DingmaomaoBJTU]

This captures input tensor shapes/types but not the node's output tensor(s). Given the branch name (add_op_size), the output shape is often the more useful "size" signal for a profiler — and an op's inputs are usually another op's outputs anyway. Was omitting onnx_outputs intentional, or worth capturing alongside inputs? Non-blocking.