Skip to content

vllm_omni.diffusion.models.ltx2.ltx2_sequence_parallel ¶

LTX2VideoToAudioParallelAttention ¶

LTX video-to-audio SP with replicated audio queries.

Video K/V arrive sequence-sharded, while the much shorter audio query is replicated. Redistribute only K/V across sequence/head dimensions, slice audio query heads locally, then gather output heads. This uses three collectives instead of standard Ulysses' four and keeps audio replicated.

enabled `property` ¶

enabled: bool

name `property` ¶

name: str

post_attention ¶

post_attention(
    attn_output: Tensor,
    ctx: ParallelAttentionContext | None,
) -> Tensor

pre_attention ¶

pre_attention(
    query: Tensor,
    key: Tensor,
    value: Tensor,
    attn_metadata: AttentionMetadata | None,
)