Driven by the widespread use of deep learning (DL) frameworks, cross-framework model conversion supports a diverse ecosystem and facilitates efficient application development. However, interestingly, existing works commonly use intermediate representations (e.g., computation graphs or APIs) for cross-framework model conversion. Due to the lack of the ability to capture operator execution details compared to source code, it is difficult for these intermediate representations to bridge semantic inconsistencies in operators. These inconsistencies change operator behavior and potentially cause poor performance and errors in converted models.

Thus, we present the first comprehensive study to analyze features and dependencies of source code related to operator semantic inconsistencies, finding that 47% of sampled operators exhibit semantic inconsistencies across DL frameworks, and the related code snippets are distributed across layers of DL frameworks without inter-layer dependencies. These findings suggest that bridging operator semantic inconsistencies layer-by-layer is feasible. Based on the findings, we propose ModelX: a source-level cross-framework model conversion approach that bridges operator semantic inconsistencies by dynamically aligning related code snippets independently within each layer, focusing on model conversion at a much finer-grained level (i.e., source code) instead of intermediate representations. The large-scale experiments on the conversion from PyTorch to Paddle show that ModelX equivalently converts 624 out of 686 sampled PyTorch operators, and yields better performance over two state-of-the-art model conversion approaches and popular large language models. Furthermore, we achieve minimal metric gaps (avg. all under 3.4%) across 52 models in the three most popular application fields (i.e., vision, text, and audio), showing that ModelX is highly robust.