fix: separate model weights from inputs in extraction output

[xjtu-L]

Summary

• Fix extractor.py to correctly separate model parameters/buffers from placeholder inputs (including SymInt symbols like s0, s1) when saving tensor metadata
• Previously, all placeholder params were passed to convert_state_and_inputs(params, []), causing SymInt placeholders to appear in weight_meta.py as Program_weight_tensor_meta_s0
• Now uses object identity (id()) to distinguish original model params/buffers from real inputs, writing weights to weight_meta.py and inputs to input_meta.py
• Updates utils.py to handle dict-style example_inputs with named tensor infos

Background

When torch.compile with FX tracing lifts model parameters and buffers into placeholder nodes, the extractor treated them identically to real model inputs. This caused:

1. SymInt symbols (s0, s1, etc.) to be written to weight_meta.py as fake weights
2. Model parameters and buffers to be written to input_meta.py instead of weight_meta.py
3. Downstream dimension generalization (_get_tensor_metas) to merge both files, propagating the redundant entries

Changes

graph_net/torch/extractor.py

• GraphExtractor.__init__: add param_buffer_ids parameter
• wrapper(): collect id() of original model's named_parameters() and named_buffers(), pass to GraphExtractor
• __call__(): split params into weights (by id match) and example_inputs, call convert_state_and_inputs(weights, example_inputs)

graph_net/torch/utils.py

• convert_state_and_inputs_impl: support example_inputs as dict (named inputs)
• save_converted_to_text: handle dict-of-tensor-infos for input_info, preserving placeholder names

Test Plan

• No-param model: weight_meta.py empty, all inputs in input_meta.py
• Nested modules: parameters correctly in weight_meta.py
• Multi-input + buffer model: parameters and buffers in weight_meta.py, real inputs in input_meta.py
• Parameter-only model (no buffer): correct split
• ParameterList model: all list params in weight_meta.py
• s0/s1 SymInt symbols never appear in weight_meta.py

[paddle-bot]

Thanks for your contribution!