[accaldwell]

Overview

• src/models/laguna.cpp previously did not load YaRN beta metadata from the model file. In the case of Laguna XS 2.1, this was falling back to a llama.cpp default 32 value for yarn_beta_fast instead of using the model file's 64. This PR loads yarn_beta_fast and yarn_beta_slow from the model file.

Additional information

Tested with Laguna XS 2.1, Q8_0 on Vulkan

• Perplexity nearly identical on three 8k chunks of the wikitext-2 test set. yarn_beta_fast 32: 10.1622 ± 0.29239 yarn_beta_fast 64: 10.1553 ± 0.29313

• Text generation quality was similar across a few temperature=0 prompts, slightly better in one case.

Requirements

• I have read and agree with the contributing guidelines
• A test is included per the contributing guidelines
• AI usage disclosure: YES. claude/Opus5.5 found this issue and codex/Sol6.1 fixed it, wrote the test, and ran the tests

[CISC]

Remove the test, otherwise OK.

[accaldwell]

Removed the test per request