Skip to content

fix: Compatibility with transformers 4.49+ and 5.0+ #137

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
xander1421 wants to merge 1 commit into
base: main
Choose a base branch
from
Open

fix: Compatibility with transformers 4.49+ and 5.0+ #137

Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
15 changes: 13 additions & 2 deletions model/dit.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -127,7 +127,16 @@ def __init__(
self.dim = dim
self.depth = depth

llama_config = LlamaConfig(hidden_size=dim, intermediate_size=dim * ff_mult, hidden_act='silu', max_position_embeddings=self.max_frames)
# Must set num_attention_heads to match DiT heads for correct head_dim in rotary embeddings
# head_dim = hidden_size // num_attention_heads, so num_attention_heads = hidden_size // dim_head
num_attention_heads = dim // dim_head
llama_config = LlamaConfig(
hidden_size=dim,
intermediate_size=dim * ff_mult,
hidden_act='silu',
max_position_embeddings=self.max_frames,
num_attention_heads=num_attention_heads, # Fix for transformers 4.49+
)
llama_config._attn_implementation = 'sdpa'
self.transformer_blocks = nn.ModuleList(
[LlamaDecoderLayer(llama_config, layer_idx=i) for i in range(depth)]
Expand Down Expand Up @@ -211,7 +220,9 @@ def forward(
)

for i, block in enumerate(self.transformer_blocks):
x, *_ = block(x, attention_mask=attention_mask, position_embeddings=rotary_embed)
# Note: In transformers 5.0+, LlamaDecoderLayer returns tensor, not tuple
# Using `x, *_ = block(...)` would unpack the tensor's first dimension
x = block(x, attention_mask=attention_mask, position_embeddings=rotary_embed)
if i < self.depth // 2:
x = x + self.text_fusion_linears[i](text_embed)

Expand Down