UPSTREAM PR #19064: Fix conversion for models that needed undo_permute = False but never declared it

[loci-dev]

Mirrored from ggml-org/llama.cpp#19064

After #18866 modify_tensors was altered to call the super function

Previously, Afmoe and CogVLM both returned the tensors directly, bypassing the super().modify_tensors() function

LlamaModel causes undo_permute to be set to True, which breaks the model by applying the Q/K permutation when the super function is called

This should fix up those cases. With this change, PPL for Trinity-Mini returns to expected levels.

This error was discovered with the help of Claude Code but tested by a human

[loci-review]

Based on the analysis, no functions were identified with meaningful performance changes between the base and target versions. The function_insights_topk tool returned empty results for both response time and throughput time metrics, indicating that the code changes between these versions did not produce measurable performance impacts in the analyzed binaries.

This suggests that the modifications were either:

• Non-performance-affecting changes (documentation, comments, refactoring)
• Changes to code paths not captured in the static analysis
• Modifications with performance impacts below the detection threshold

Conclusion: No significant performance regression or improvement was detected between the two versions.

See the complete breakdown in Version Insights
Have questions? Tag @loci-dev to ask about this PR.