Skip loading external weight data during static analysis.

[chinazhangchao]

Problem

Running winml analyze on large models with external data (e.g., Qwen3-8B with a 30.5 GB .data sidecar) causes the process to consume all available memory and disk, hanging indefinitely. The analyzer called onnx.load(path, load_external_data=True), loading the entire weight file into RAM despite never inspecting weight values.

Root Cause

The static analyzer only needs graph structure (operator types, shapes, connectivity, and small embedded constants) to perform op-support checks. Three call sites were loading or attempting to access the full weight tensors unnecessarily:

1. ONNXStaticAnalyzer.analyze() — explicitly passed load_external_data=True
2. ONNXLoader.load() — called bare onnx.load() which defaults to load_external_data=True
3. RuntimeCheckerQuery — called numpy_helper.to_array() on every initializer and embedded full TensorProtos into single-node models

Changes

File	Change
analyzer.py	load_external_data=True → False
onnx_loader.py	Add explicit load_external_data=False
runtime_checker_query.py	For external-data initializers: extract shape from dims instead of to_array(), and emit graph inputs instead of embedding empty tensors in single-node models

Impact

• Models with external data (Qwen3-8B, Llama, etc.) now use ~MB instead of ~30+ GB RAM
• No behavior change for models with inline weights (the data_location != EXTERNAL path is unchanged)
• 849 unit tests pass, 0 failures

Performance (Improve 39.65%)

Model	Operators	Time
Qwen/Qwen3-8B	5333	57.36s
dbmdz/bert-large-cased-finetuned-conll03-english	2663	28.395s

[DingmaomaoBJTU]

Overall good fix for a real and impactful problem — static analysis never needs multi-GB weight tensors, and the three-site fix is comprehensive. A few items to address before merging.

[DingmaomaoBJTU]

Good work addressing the previous round of feedback — the is_constant=False fix, test coverage additions, and the raw_data comment all look solid. The three-site fix is correct and comprehensive. A few new items for this round:

Non-inline note — _collect_node_tags / ALL_INPUTS_CONSTANT:
External-data initializers are still in self.initializers, so a node whose inputs are all unloaded external weights would be tagged ALL_INPUTS_CONSTANT even though weight data is unavailable. This is informational-only and doesn't affect runtime check results, but could confuse future debugging. Consider filtering external-data initializers without loaded data in that check.