Class ov::pass::DisableBF16CompForLtxVideoRopePattern

class DisableBF16CompForLtxVideoRopePattern : public ov::pass::MatcherPass

Keeps the decomposed LTX-Video RoPE angle chain (Multiply -> Add -> Transpose -> Reshape -> Sin/Cos) in f32: the angles reach ~1.6e4 rad, which bf16 quantizes in steps larger than 2*pi.

The CPU plugin registers this pass only under bf16 enforcement. The fused RoPE variant is handled on GPU by IncreasePositionIdsPrecisionForLtxVideo.