UPSTREAM PR #18479: support youtu-vl model

[loci-dev]

Mirrored from ggml-org/llama.cpp#18479

Make sure to read the contributing guidelines before submitting a PR

[loci-review]

Explore the complete analysis inside the Version Insights

I've successfully retrieved the summary report for your project. The report shows that Pull Request #755 for the auroralabs-loci/llama.cpp repository has significant performance regressions:

Key Highlights:

• Top Issue: STL container operations (tree and vector iterators) show dramatic performance degradation, with response times increasing by 43% to 230%
• Most Affected Functions:
  • end and begin functions in stl_tree.h (+230% and +220% response time)
  • Vector iterator operations (+215% response time)
• Affected Binaries: Primarily libllama.so and libmtmd.so

The report suggests investigating changes to data structure usage patterns and iterator efficiency, as this PR appears to introduce substantial performance regressions that should be addressed before merging.

Would you like more detailed information about any specific aspect of this report?

[loci-review]

Explore the complete analysis inside the Version Insights

I've successfully retrieved the summary report for your llama.cpp project (PR #755). The report shows the top 10 functions with the most significant performance changes between the base and target versions.

Key Highlights:

• Major Regressions: Several STL vector begin() methods show dramatic response time increases of +215.7%, and the clip_hparams destructor shows +92.3% increase
• Improvements: The begin() method for vector<llm_symbol> improved by -68.3%, and ma_decoder_config_init improved by -8.1%
• Overall: The PR shows mixed performance impacts with most functions experiencing regressions

Would you like more detailed analysis on any specific function or aspect of this performance report?