Added Preview mode and support Sky Reels v2 Diffusion Forcing
This commit is contained in:
@@ -1075,13 +1075,13 @@ def sageattn_qk_int8_pv_fp8_cuda_sm90(
|
||||
q_int8, q_scale, k_int8, k_scale = per_thread_int8_triton(q, k, km, tensor_layout=tensor_layout, BLKQ=64, WARPQ=16, BLKK=128, WARPK=128)
|
||||
|
||||
q_size = q.size()
|
||||
kv_len = k.size(seq_dim)
|
||||
q_device = q.device
|
||||
del q,k
|
||||
|
||||
|
||||
# pad v to multiple of 128
|
||||
# TODO: modify per_channel_fp8 kernel to handle this
|
||||
kv_len = k.size(seq_dim)
|
||||
v_pad_len = 128 - (kv_len % 128) if kv_len % 128 != 0 else 0
|
||||
if v_pad_len > 0:
|
||||
if tensor_layout == "HND":
|
||||
|
||||
Reference in New Issue
Block a user