UPSTREAM PR #18548: graph : reduce topology branching

[loci-dev]

Mirrored from ggml-org/llama.cpp#18548

ref #18547

Reduce the amount of graph topology changes when switching between token and embedding inputs:

// before
if (ubatch.token) {
    inpL = ggml_scale(ctx0, inpL, sqrtf(n_embd));
    cb(inpL, "inp_scaled", -1);
}

// after
inpL = ggml_scale(ctx0, inpL, ubatch.token ? sqrtf(n_embd) : 1.0f);
cb(inpL, "inp_scaled", -1);

[loci-review]

Explore the complete analysis inside the Version Insights

Perfect! I've retrieved the summary report for your project. Here are the key findings:

Summary Report for llama.cpp PR #790

Project Information:

• Repository: llama.cpp (owner: auroralabs-loci)
• Pull Request: UPSTREAM PR #18548: graph : reduce topology branching #790
• Comparing versions between base and target

Performance Analysis Results:

✅ No Significant Performance Regressions Detected

The analysis shows that Pull Request #790 maintains performance parity with the base version:

• Response Time Changes: No modified functions with performance changes > 2%
• Throughput Time Changes: No modified functions with performance changes > 2%

Conclusion:
This is a positive result indicating that the changes in PR #790 are performance-neutral. The modifications don't negatively impact the performance of the llama.cpp project, suggesting the changes are likely focused on functionality improvements, bug fixes, or refactoring while maintaining the existing performance characteristics.