attn-rotation for native llama quants

[erazortt]

it would be helpfull if the rotation of the native llama quants would be available.
Especially so for the Q8-turboX KV combination. This would then improve in quality.

[TheTom]

@erazortt thanks for flagging this. you were right.

dug into the code and found we were defaulting LLAMA_ATTN_ROT_DISABLE=true in llama_kv_cache::cparams constructor (src/llama-kv-cache.cpp:445), which globally suppressed the upstream activation-rotation pass from ml-explore/llama.cpp#21038. the original rationale was "TurboQuant uses kernel-level WHT" — true for symmetric turbo configs, wrong for asymmetric (q8_0-K + turbo*-V) where K stays unrotated and loses the Q8 quality boost.

empirically confirmed on Qwen3.5-2B Q8_0, q8_0-K + turbo4-V cache, wikitext-2 16 chunks @ ctx=2048:

config	attn_rot_k	attn_rot_v	PPL
pre-fix default (rotation off)	0	0	10.9170 ± 0.233
fix default (rotation on)	1	1	10.8819 ± 0.235

-0.32% PPL, 15 of 16 chunks favorable. small but consistent.

PR up at TheTom/llama-cpp-turboquant#111. could you pull that branch (fix/enable-attn-rot-by-default) and run your test setup on whichever model showed the gap most clearly? particularly interested in larger model classes where the delta might widen.

LLAMA_ATTN_ROT_DISABLE=1 is still available as a global override if a model hits graph-node hash overflow (Phi-4 was the prior failure case).