Fix Slice and Split derive_properties for dynamic axes

[vortex-captain]

Summary

• Fix Slice derive_properties crash when data_shape contains symbolic dimension names (e.g. "time" from dynamic axes). Filters to fixed-shape axes only when computing starts_equal_shape and slice_all.
• Fix Split derive_properties to use n_outputs from caller when split_value and attr_num_outputs are both absent.

[DingmaomaoBJTU]

Code review

Found 1 issue:

1. Pattern path injects incorrect n_outputs for variadic-output operators like Split

The node path correctly uses len(node.output) (the actual number of output edges on the matched node). The pattern path uses len(pattern_match.skeleton_match_result.pattern.get_schema().outputs), which returns the number of schema-defined output parameters — not the number of actual output edges. For the Split operator, the ONNX schema defines a single variadic output parameter, so get_schema().outputs always has length 1 regardless of how many ways the tensor is split. Any Split node matched via a pattern will receive n_outputs = 1, causing SplitInputGenerator.derive_properties to derive num_outputs = 1 instead of the real value.

https://github.com/microsoft/ModelKit/blob/1114c0c3da63e6970a1880ddff273a6f1c6bfeee/src/winml/modelkit/analyze/core/runtime_checker_query.py#L820-L824

The node path (used in get_query_conditions_for_node) may be the right model here — something like the number of matched output edges rather than the schema parameter count.

🤖 Generated with Claude Code

_{- If this code review was useful, please react with 👍. Otherwise, react with 👎.}

[vortex-captain]

Code review

Found 1 issue:

1. Pattern path injects incorrect n_outputs for variadic-output operators like Split

The node path correctly uses len(node.output) (the actual number of output edges on the matched node). The pattern path uses len(pattern_match.skeleton_match_result.pattern.get_schema().outputs), which returns the number of schema-defined output parameters — not the number of actual output edges. For the Split operator, the ONNX schema defines a single variadic output parameter, so get_schema().outputs always has length 1 regardless of how many ways the tensor is split. Any Split node matched via a pattern will receive n_outputs = 1, causing SplitInputGenerator.derive_properties to derive num_outputs = 1 instead of the real value.

https://github.com/microsoft/ModelKit/blob/1114c0c3da63e6970a1880ddff273a6f1c6bfeee/src/winml/modelkit/analyze/core/runtime_checker_query.py#L820-L824

The node path (used in get_query_conditions_for_node) may be the right model here — something like the number of matched output edges rather than the schema parameter count.

🤖 Generated with Claude Code

• If this code review was useful, please react with 👍. Otherwise, react with 👎.

Matching subgraph patterns with variadic outputs is out of scope of pattern matcher, as it complicates things tremendously and has no practical use cases for now. The current conditions["n_outputs"] works properly in the current scope.