Move RoPE parameters into model config
#2
by User01110 - opened
Expose the trained split-half RoPE settings in config.json, source rope_theta from ACSwiGLUConfig, validate the fixed no-scaling layout, and rebuild deterministic RoPE caches instead of storing them as checkpoint buffers.
User01110 changed pull request status to open
Validated by Python compilation and JSON parsing. Merging the configuration-only RoPE cleanup; model weights and trained numerical behavior remain unchanged.
User01110 changed pull request status to merged