PyTorch SDPA's is_causal=True with Q != K uses tril(diagonal=0) which is
top-left aligned. For a single decode step (Q=1) against a long ViT+text
cache (K=574+), only q_0 -> k_0 is unmasked — producing degenerate output
("TheTheThe..." token loops).
Bypass DiffusionAttention (whose _maybe_reshape_attn_mask only handles
2-D (B, K) masks) and call torch.nn.functional.scaled_dot_product_attention
directly with an explicit 4-D (1, 1, Q, K) bottom-right-aligned causal mask
when cache_len > 0. Also cast q/k to fp32 before RMSNorm for precision.
Verified all 7 tasks on Lance:
- x2t_image: "No" (valid VQA answer)
- x2t_video: "The purple sphere ... right direc..." (valid VQA answer)
- t2i / i2v / t2v / image_edit / video_edit all produce expected media.
Signed-off-by: lishunyang12 <lishunyang12@163.com>
lishunyang12
18252b9d8d