feat(wan): SP communication-computation overlap for Wan2.2 I2V.

[nie-linfeng]

背景

Wan2.2 I2V 模型在 SP (Sequence Parallel) 模式下，attention 层的 all2all 通信与计算串行执行，导致通信开销无法被掩盖。

优化方案

通过 async all2all 将通信与计算 overlap，分两个层面优化：

1. Self-Attention (forward_async)

基线流程（串行）：

Q matmul -> norm Q -> Q all2all
K matmul -> norm K -> K all2all
V matmul -> V all2all (to_v_ sp_before)
reshape -> RoPE
attention
O all2all -> to_out matmul (to_out_ sp_after)

CMO 流程（async overlap）：

Q matmul -> norm Q -> async Q all2all
K matmul -> norm K -> async K all2all
V matmul -> async V all2all
Wait Q/K -> reshape -> RoPE (V all2all overlap)
Wait V
attention
async O all2all
return handler (wait O -> to_out matmul)

关键点：

• Q/K/V 三个 all2all 以 async 模式发起
• V all2all 与 Q/K reshape+RoPE overlap（部分掩盖）
• O all2all async，handler 返回给调用方

2. Cross-Attention (forward_cross)

基线流程（串行）：

Q matmul -> norm Q -> Q all2all
K matmul -> norm K -> sp_slice_heads
V matmul -> sp_slice_heads
attention
O all2all -> to_out matmul

CMO 流程（async overlap）：

Q matmul -> norm Q -> async Q all2all
K matmul -> norm K -> sp_slice_heads (overlap with Q all2all)
V matmul -> sp_slice_heads
img K/V matmul -> sp_slice_heads
Wait Q
attention
O all2all -> wait -> to_out matmul

关键点：

• Q all2all async 后立即计算 K/V(text+img) matmul+norm+slice
• K/V matmul 计算掩盖 Q all2all 通信

3. 配置改动

• CMO 启用时 to_v_ 从 TP+SP(sp_before) 改为 TP-only
• CMO 启用时 to_out_ 从 TP+SP(sp_after) 改为 TP-only
• wan_use_sp_overlap() 控制

改动文件

• transformer_wan.h: 新增 WanAttnProcessorCMO(forward_async + forward_cross)
• dit_parallel_linear.h: TP+SP combined linear 支持

性能数据

配置	稳定平均	vs 基线
基线 (main)	~106.4s	--
CMO	~103.1s	-3.1%

测试条件: 8 dieNPU, tp1, sp4, vae_size 4 ;40 steps

Profiling

优化后：
self-attn之前qkv的all-to-all

cross-attn之前q的all-to-all