github/FramePack
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

prepare for sync

Browse Source
This commit is contained in:
layerdiffusion 2025-05-02 07:09:16 -07:00
parent 6da55e8f87
commit ef4d1ddda4
1 changed files with 23 additions and 20 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
diffusers_helper/models/hunyuan_video_packed.py
View File

@ -122,17 +122,21 @@ def attn_varlen_func(q, k, v, cu_seqlens_q, cu_seqlens_kv, max_seqlen_q, max_seq
x = torch.nn.functional.scaled_dot_product_attention(q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)).transpose(1, 2) x = torch.nn.functional.scaled_dot_product_attention(q.transpose(1, 2), k.transpose(1, 2), v.transpose(1, 2)).transpose(1, 2)
return x return x
batch_size = q.shape[0] B, L, H, C = q.shape
q = q.view(q.shape[0] * q.shape[1], *q.shape[2:])
k = k.view(k.shape[0] * k.shape[1], *k.shape[2:]) q = q.flatten(0, 1)
v = v.view(v.shape[0] * v.shape[1], *v.shape[2:]) k = k.flatten(0, 1)
v = v.flatten(0, 1)
if sageattn_varlen is not None: if sageattn_varlen is not None:
x = sageattn_varlen(q, k, v, cu_seqlens_q, cu_seqlens_kv, max_seqlen_q, max_seqlen_kv) x = sageattn_varlen(q, k, v, cu_seqlens_q, cu_seqlens_kv, max_seqlen_q, max_seqlen_kv)
elif flash_attn_varlen_func is not None: elif flash_attn_varlen_func is not None:
x = flash_attn_varlen_func(q, k, v, cu_seqlens_q, cu_seqlens_kv, max_seqlen_q, max_seqlen_kv) x = flash_attn_varlen_func(q, k, v, cu_seqlens_q, cu_seqlens_kv, max_seqlen_q, max_seqlen_kv)
else: else:
raise NotImplementedError('No Attn Installed!') raise NotImplementedError('No Attn Installed!')
x = x.view(batch_size, max_seqlen_q, *x.shape[2:])
x = x.unflatten(0, (B, L))
return x return x
@ -926,7 +930,6 @@ class HunyuanVideoTransformer3DModelPacked(ModelMixin, ConfigMixin, PeftAdapterM
encoder_hidden_states = torch.cat([extra_encoder_hidden_states, encoder_hidden_states], dim=1) encoder_hidden_states = torch.cat([extra_encoder_hidden_states, encoder_hidden_states], dim=1)
encoder_attention_mask = torch.cat([extra_attention_mask, encoder_attention_mask], dim=1) encoder_attention_mask = torch.cat([extra_attention_mask, encoder_attention_mask], dim=1)
with torch.no_grad():
if batch_size == 1: if batch_size == 1:
# When batch size is 1, we do not need any masks or var-len funcs since cropping is mathematically same to what we want # When batch size is 1, we do not need any masks or var-len funcs since cropping is mathematically same to what we want
# If they are not same, then their impls are wrong. Ours are always the correct one. # If they are not same, then their impls are wrong. Ours are always the correct one.