Skip to content

[WIP] Refactor Model Design #13794

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
DN6 wants to merge 9 commits into
base: main
Choose a base branch
from
Open

[WIP] Refactor Model Design #13794

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
4c90ebc
Refactor model design: metadata system + Flux port to package with We…
DN6 Jul 1, 2026
c90e57f
clean up
DN6 Jul 1, 2026
14d0722
clean up
DN6 Jul 1, 2026
1953e6b
clean up
DN6 Jul 1, 2026
698e7f3
fix
DN6 Jul 1, 2026
f514d7a
clean up
DN6 Jul 1, 2026
4763076
clean up
DN6 Jul 1, 2026
7f58967
Merge remote-tracking branch 'origin/main' into refactor-model-metadata
DN6 Oct 29, 2026
7c88feb
Merge remote-tracking branch 'origin/main' into refactor-model-metadata
DN6 Oct 30, 2026
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
4 changes: 2 additions & 2 deletions src/diffusers/hooks/_helpers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -108,8 +108,8 @@ def _register(cls):

def _register_attention_processors_metadata():
from ..models.attention_processor import AttnProcessor2_0
from ..models.transformers.flux import FluxAttnProcessor
from ..models.transformers.transformer_cogview4 import CogView4AttnProcessor
from ..models.transformers.transformer_flux import FluxAttnProcessor
from ..models.transformers.transformer_hunyuanimage import HunyuanImageAttnProcessor
from ..models.transformers.transformer_qwenimage import QwenDoubleStreamAttnProcessor2_0
from ..models.transformers.transformer_wan import WanAttnProcessor2_0
Expand Down Expand Up @@ -173,10 +173,10 @@ def _register_attention_processors_metadata():
def _register_transformer_blocks_metadata():
from ..models.attention import BasicTransformerBlock, JointTransformerBlock
from ..models.transformers.cogvideox_transformer_3d import CogVideoXBlock
from ..models.transformers.flux import FluxSingleTransformerBlock, FluxTransformerBlock
from ..models.transformers.transformer_bria import BriaTransformerBlock
from ..models.transformers.transformer_cogview4 import CogView4TransformerBlock
from ..models.transformers.transformer_cosmos3 import Cosmos3VLTextMoTDecoderLayer
from ..models.transformers.transformer_flux import FluxSingleTransformerBlock, FluxTransformerBlock
from ..models.transformers.transformer_flux2 import Flux2SingleTransformerBlock, Flux2TransformerBlock
from ..models.transformers.transformer_hunyuan_video import (
HunyuanVideoSingleTransformerBlock,
Expand Down
5 changes: 0 additions & 5 deletions src/diffusers/loaders/single_file_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,7 +40,6 @@
convert_ernie_image_transformer_checkpoint_to_diffusers,
convert_flux2_transformer_checkpoint_to_diffusers,
convert_flux3_transformer_checkpoint_to_diffusers,
convert_flux_transformer_checkpoint_to_diffusers,
convert_hidream_transformer_to_diffusers,
convert_hunyuan_video_transformer_to_diffusers,
convert_krea2_transformer_checkpoint_to_diffusers,
Expand Down Expand Up @@ -116,10 +115,6 @@
"SparseControlNetModel": {
"checkpoint_mapping_fn": convert_animatediff_checkpoint_to_diffusers,
},
"FluxTransformer2DModel": {
"checkpoint_mapping_fn": convert_flux_transformer_checkpoint_to_diffusers,
"default_subfolder": "transformer",
},
"ChromaTransformer2DModel": {
"checkpoint_mapping_fn": convert_chroma_transformer_checkpoint_to_diffusers,
"default_subfolder": "transformer",
Expand Down
197 changes: 0 additions & 197 deletions src/diffusers/loaders/single_file_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2291,203 +2291,6 @@ def convert_animatediff_checkpoint_to_diffusers(checkpoint, **kwargs):
return converted_state_dict


def convert_flux_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
converted_state_dict = {}
keys = list(checkpoint.keys())

for k in keys:
if "model.diffusion_model." in k:
checkpoint[k.replace("model.diffusion_model.", "")] = checkpoint.pop(k)

num_layers = list(set(int(k.split(".", 2)[1]) for k in checkpoint if "double_blocks." in k))[-1] + 1 # noqa: C401
num_single_layers = list(set(int(k.split(".", 2)[1]) for k in checkpoint if "single_blocks." in k))[-1] + 1 # noqa: C401
mlp_ratio = 4.0
inner_dim = 3072

# in SD3 original implementation of AdaLayerNormContinuous, it split linear projection output into shift, scale;
# while in diffusers it split into scale, shift. Here we swap the linear projection weights in order to be able to use diffusers implementation
def swap_scale_shift(weight):
shift, scale = weight.chunk(2, dim=0)
new_weight = torch.cat([scale, shift], dim=0)
return new_weight

## time_text_embed.timestep_embedder <- time_in
converted_state_dict["time_text_embed.timestep_embedder.linear_1.weight"] = checkpoint.pop(
"time_in.in_layer.weight"
)
converted_state_dict["time_text_embed.timestep_embedder.linear_1.bias"] = checkpoint.pop("time_in.in_layer.bias")
converted_state_dict["time_text_embed.timestep_embedder.linear_2.weight"] = checkpoint.pop(
"time_in.out_layer.weight"
)
converted_state_dict["time_text_embed.timestep_embedder.linear_2.bias"] = checkpoint.pop("time_in.out_layer.bias")

## time_text_embed.text_embedder <- vector_in
converted_state_dict["time_text_embed.text_embedder.linear_1.weight"] = checkpoint.pop("vector_in.in_layer.weight")
converted_state_dict["time_text_embed.text_embedder.linear_1.bias"] = checkpoint.pop("vector_in.in_layer.bias")
converted_state_dict["time_text_embed.text_embedder.linear_2.weight"] = checkpoint.pop(
"vector_in.out_layer.weight"
)
converted_state_dict["time_text_embed.text_embedder.linear_2.bias"] = checkpoint.pop("vector_in.out_layer.bias")

# guidance
has_guidance = any("guidance" in k for k in checkpoint)
if has_guidance:
converted_state_dict["time_text_embed.guidance_embedder.linear_1.weight"] = checkpoint.pop(
"guidance_in.in_layer.weight"
)
converted_state_dict["time_text_embed.guidance_embedder.linear_1.bias"] = checkpoint.pop(
"guidance_in.in_layer.bias"
)
converted_state_dict["time_text_embed.guidance_embedder.linear_2.weight"] = checkpoint.pop(
"guidance_in.out_layer.weight"
)
converted_state_dict["time_text_embed.guidance_embedder.linear_2.bias"] = checkpoint.pop(
"guidance_in.out_layer.bias"
)

# context_embedder
converted_state_dict["context_embedder.weight"] = checkpoint.pop("txt_in.weight")
converted_state_dict["context_embedder.bias"] = checkpoint.pop("txt_in.bias")

# x_embedder
converted_state_dict["x_embedder.weight"] = checkpoint.pop("img_in.weight")
converted_state_dict["x_embedder.bias"] = checkpoint.pop("img_in.bias")

# double transformer blocks
for i in range(num_layers):
block_prefix = f"transformer_blocks.{i}."
# norms.
## norm1
converted_state_dict[f"{block_prefix}norm1.linear.weight"] = checkpoint.pop(
f"double_blocks.{i}.img_mod.lin.weight"
)
converted_state_dict[f"{block_prefix}norm1.linear.bias"] = checkpoint.pop(
f"double_blocks.{i}.img_mod.lin.bias"
)
## norm1_context
converted_state_dict[f"{block_prefix}norm1_context.linear.weight"] = checkpoint.pop(
f"double_blocks.{i}.txt_mod.lin.weight"
)
converted_state_dict[f"{block_prefix}norm1_context.linear.bias"] = checkpoint.pop(
f"double_blocks.{i}.txt_mod.lin.bias"
)
# Q, K, V
sample_q, sample_k, sample_v = torch.chunk(checkpoint.pop(f"double_blocks.{i}.img_attn.qkv.weight"), 3, dim=0)
context_q, context_k, context_v = torch.chunk(
checkpoint.pop(f"double_blocks.{i}.txt_attn.qkv.weight"), 3, dim=0
)
sample_q_bias, sample_k_bias, sample_v_bias = torch.chunk(
checkpoint.pop(f"double_blocks.{i}.img_attn.qkv.bias"), 3, dim=0
)
context_q_bias, context_k_bias, context_v_bias = torch.chunk(
checkpoint.pop(f"double_blocks.{i}.txt_attn.qkv.bias"), 3, dim=0
)
converted_state_dict[f"{block_prefix}attn.to_q.weight"] = torch.cat([sample_q])
converted_state_dict[f"{block_prefix}attn.to_q.bias"] = torch.cat([sample_q_bias])
converted_state_dict[f"{block_prefix}attn.to_k.weight"] = torch.cat([sample_k])
converted_state_dict[f"{block_prefix}attn.to_k.bias"] = torch.cat([sample_k_bias])
converted_state_dict[f"{block_prefix}attn.to_v.weight"] = torch.cat([sample_v])
converted_state_dict[f"{block_prefix}attn.to_v.bias"] = torch.cat([sample_v_bias])
converted_state_dict[f"{block_prefix}attn.add_q_proj.weight"] = torch.cat([context_q])
converted_state_dict[f"{block_prefix}attn.add_q_proj.bias"] = torch.cat([context_q_bias])
converted_state_dict[f"{block_prefix}attn.add_k_proj.weight"] = torch.cat([context_k])
converted_state_dict[f"{block_prefix}attn.add_k_proj.bias"] = torch.cat([context_k_bias])
converted_state_dict[f"{block_prefix}attn.add_v_proj.weight"] = torch.cat([context_v])
converted_state_dict[f"{block_prefix}attn.add_v_proj.bias"] = torch.cat([context_v_bias])
# qk_norm
converted_state_dict[f"{block_prefix}attn.norm_q.weight"] = checkpoint.pop(
f"double_blocks.{i}.img_attn.norm.query_norm.scale"
)
converted_state_dict[f"{block_prefix}attn.norm_k.weight"] = checkpoint.pop(
f"double_blocks.{i}.img_attn.norm.key_norm.scale"
)
converted_state_dict[f"{block_prefix}attn.norm_added_q.weight"] = checkpoint.pop(
f"double_blocks.{i}.txt_attn.norm.query_norm.scale"
)
converted_state_dict[f"{block_prefix}attn.norm_added_k.weight"] = checkpoint.pop(
f"double_blocks.{i}.txt_attn.norm.key_norm.scale"
)
# ff img_mlp
converted_state_dict[f"{block_prefix}ff.net.0.proj.weight"] = checkpoint.pop(
f"double_blocks.{i}.img_mlp.0.weight"
)
converted_state_dict[f"{block_prefix}ff.net.0.proj.bias"] = checkpoint.pop(f"double_blocks.{i}.img_mlp.0.bias")
converted_state_dict[f"{block_prefix}ff.net.2.weight"] = checkpoint.pop(f"double_blocks.{i}.img_mlp.2.weight")
converted_state_dict[f"{block_prefix}ff.net.2.bias"] = checkpoint.pop(f"double_blocks.{i}.img_mlp.2.bias")
converted_state_dict[f"{block_prefix}ff_context.net.0.proj.weight"] = checkpoint.pop(
f"double_blocks.{i}.txt_mlp.0.weight"
)
converted_state_dict[f"{block_prefix}ff_context.net.0.proj.bias"] = checkpoint.pop(
f"double_blocks.{i}.txt_mlp.0.bias"
)
converted_state_dict[f"{block_prefix}ff_context.net.2.weight"] = checkpoint.pop(
f"double_blocks.{i}.txt_mlp.2.weight"
)
converted_state_dict[f"{block_prefix}ff_context.net.2.bias"] = checkpoint.pop(
f"double_blocks.{i}.txt_mlp.2.bias"
)
# output projections.
converted_state_dict[f"{block_prefix}attn.to_out.0.weight"] = checkpoint.pop(
f"double_blocks.{i}.img_attn.proj.weight"
)
converted_state_dict[f"{block_prefix}attn.to_out.0.bias"] = checkpoint.pop(
f"double_blocks.{i}.img_attn.proj.bias"
)
converted_state_dict[f"{block_prefix}attn.to_add_out.weight"] = checkpoint.pop(
f"double_blocks.{i}.txt_attn.proj.weight"
)
converted_state_dict[f"{block_prefix}attn.to_add_out.bias"] = checkpoint.pop(
f"double_blocks.{i}.txt_attn.proj.bias"
)

# single transformer blocks
for i in range(num_single_layers):
block_prefix = f"single_transformer_blocks.{i}."
# norm.linear <- single_blocks.0.modulation.lin
converted_state_dict[f"{block_prefix}norm.linear.weight"] = checkpoint.pop(
f"single_blocks.{i}.modulation.lin.weight"
)
converted_state_dict[f"{block_prefix}norm.linear.bias"] = checkpoint.pop(
f"single_blocks.{i}.modulation.lin.bias"
)
# Q, K, V, mlp
mlp_hidden_dim = int(inner_dim * mlp_ratio)
split_size = (inner_dim, inner_dim, inner_dim, mlp_hidden_dim)
q, k, v, mlp = torch.split(checkpoint.pop(f"single_blocks.{i}.linear1.weight"), split_size, dim=0)
q_bias, k_bias, v_bias, mlp_bias = torch.split(
checkpoint.pop(f"single_blocks.{i}.linear1.bias"), split_size, dim=0
)
converted_state_dict[f"{block_prefix}attn.to_q.weight"] = torch.cat([q])
converted_state_dict[f"{block_prefix}attn.to_q.bias"] = torch.cat([q_bias])
converted_state_dict[f"{block_prefix}attn.to_k.weight"] = torch.cat([k])
converted_state_dict[f"{block_prefix}attn.to_k.bias"] = torch.cat([k_bias])
converted_state_dict[f"{block_prefix}attn.to_v.weight"] = torch.cat([v])
converted_state_dict[f"{block_prefix}attn.to_v.bias"] = torch.cat([v_bias])
converted_state_dict[f"{block_prefix}proj_mlp.weight"] = torch.cat([mlp])
converted_state_dict[f"{block_prefix}proj_mlp.bias"] = torch.cat([mlp_bias])
# qk norm
converted_state_dict[f"{block_prefix}attn.norm_q.weight"] = checkpoint.pop(
f"single_blocks.{i}.norm.query_norm.scale"
)
converted_state_dict[f"{block_prefix}attn.norm_k.weight"] = checkpoint.pop(
f"single_blocks.{i}.norm.key_norm.scale"
)
# output projections.
converted_state_dict[f"{block_prefix}proj_out.weight"] = checkpoint.pop(f"single_blocks.{i}.linear2.weight")
converted_state_dict[f"{block_prefix}proj_out.bias"] = checkpoint.pop(f"single_blocks.{i}.linear2.bias")

converted_state_dict["proj_out.weight"] = checkpoint.pop("final_layer.linear.weight")
converted_state_dict["proj_out.bias"] = checkpoint.pop("final_layer.linear.bias")
converted_state_dict["norm_out.linear.weight"] = swap_scale_shift(
checkpoint.pop("final_layer.adaLN_modulation.1.weight")
)
converted_state_dict["norm_out.linear.bias"] = swap_scale_shift(
checkpoint.pop("final_layer.adaLN_modulation.1.bias")
)

return converted_state_dict


def convert_ltx_transformer_checkpoint_to_diffusers(checkpoint, **kwargs):
converted_state_dict = {key: checkpoint.pop(key) for key in list(checkpoint.keys()) if "vae" not in key}

Expand Down
Loading
Loading