Pull requests / #2967
#2967 feat: experimental RoPE to YaRN 4x saved-cache migration
open · @CC-David-CC · 0 commentaires · Sur GitHub
AMD / HIPNVIDIA / CUDADocumentation
Description
Opt-in saved-session migration from ordinary RoPE to approximate YaRN 4x history, including MTP draft keys. Requires FP16 KV, text-only, one GPU, `--experimental-rope-yarn4-cache` and `STRATA_ROPE_TABLE=1`. Usage: `docs/ROPE_CACHE_MIGRATION.md`. Minimal replacement for #1698, refreshed onto current main. Defaults unchanged. Validated: CUDA/HIP migration and save/reload, native and slot API tests, SYCL build, and default-off token identity against main.
Sur le site
Liens install, modèles, releases.