mirror of
https://github.com/modelscope/DiffSynth-Studio.git
synced 2026-03-21 16:18:13 +00:00
flux
This commit is contained in:
yjy415
@@ -1,9 +1,62 @@
|
||||
import torch
|
||||
from einops import rearrange, repeat
|
||||
from .flux_dit import RoPEEmbedding, TimestepEmbeddings, FluxJointTransformerBlock, FluxSingleTransformerBlock, RMSNorm
|
||||
from .utils import hash_state_dict_keys, init_weights_on_device
|
||||
# from .utils import hash_state_dict_keys, init_weights_on_device
|
||||
from contextlib import contextmanager
|
||||
|
||||
def hash_state_dict_keys(state_dict, with_shape=True):
|
||||
keys_str = convert_state_dict_keys_to_single_str(state_dict, with_shape=with_shape)
|
||||
keys_str = keys_str.encode(encoding="UTF-8")
|
||||
return hashlib.md5(keys_str).hexdigest()
|
||||
|
||||
@contextmanager
|
||||
def init_weights_on_device(device = torch.device("meta"), include_buffers :bool = False):
|
||||
|
||||
old_register_parameter = torch.nn.Module.register_parameter
|
||||
if include_buffers:
|
||||
old_register_buffer = torch.nn.Module.register_buffer
|
||||
|
||||
def register_empty_parameter(module, name, param):
|
||||
old_register_parameter(module, name, param)
|
||||
if param is not None:
|
||||
param_cls = type(module._parameters[name])
|
||||
kwargs = module._parameters[name].__dict__
|
||||
kwargs["requires_grad"] = param.requires_grad
|
||||
module._parameters[name] = param_cls(module._parameters[name].to(device), **kwargs)
|
||||
|
||||
def register_empty_buffer(module, name, buffer, persistent=True):
|
||||
old_register_buffer(module, name, buffer, persistent=persistent)
|
||||
if buffer is not None:
|
||||
module._buffers[name] = module._buffers[name].to(device)
|
||||
|
||||
def patch_tensor_constructor(fn):
|
||||
def wrapper(*args, **kwargs):
|
||||
kwargs["device"] = device
|
||||
return fn(*args, **kwargs)
|
||||
|
||||
return wrapper
|
||||
|
||||
if include_buffers:
|
||||
tensor_constructors_to_patch = {
|
||||
torch_function_name: getattr(torch, torch_function_name)
|
||||
for torch_function_name in ["empty", "zeros", "ones", "full"]
|
||||
}
|
||||
else:
|
||||
tensor_constructors_to_patch = {}
|
||||
|
||||
try:
|
||||
torch.nn.Module.register_parameter = register_empty_parameter
|
||||
if include_buffers:
|
||||
torch.nn.Module.register_buffer = register_empty_buffer
|
||||
for torch_function_name in tensor_constructors_to_patch.keys():
|
||||
setattr(torch, torch_function_name, patch_tensor_constructor(getattr(torch, torch_function_name)))
|
||||
yield
|
||||
finally:
|
||||
torch.nn.Module.register_parameter = old_register_parameter
|
||||
if include_buffers:
|
||||
torch.nn.Module.register_buffer = old_register_buffer
|
||||
for torch_function_name, old_torch_function in tensor_constructors_to_patch.items():
|
||||
setattr(torch, torch_function_name, old_torch_function)
|
||||
|
||||
class FluxControlNet(torch.nn.Module):
|
||||
def __init__(self, disable_guidance_embedder=False, num_joint_blocks=5, num_single_blocks=10, num_mode=0, mode_dict={}, additional_input_dim=0):
|
||||
@@ -102,9 +155,9 @@ class FluxControlNet(torch.nn.Module):
|
||||
return controlnet_res_stack, controlnet_single_res_stack
|
||||
|
||||
|
||||
@staticmethod
|
||||
def state_dict_converter():
|
||||
return FluxControlNetStateDictConverter()
|
||||
# @staticmethod
|
||||
# def state_dict_converter():
|
||||
# return FluxControlNetStateDictConverter()
|
||||
|
||||
def quantize(self):
|
||||
def cast_to(weight, dtype=None, device=None, copy=False):
|
||||
|
||||
@@ -1,5 +1,415 @@
|
||||
import torch
|
||||
from .sd_text_encoder import CLIPEncoderLayer
|
||||
from einops import rearrange
|
||||
|
||||
|
||||
def low_version_attention(query, key, value, attn_bias=None):
|
||||
scale = 1 / query.shape[-1] ** 0.5
|
||||
query = query * scale
|
||||
attn = torch.matmul(query, key.transpose(-2, -1))
|
||||
if attn_bias is not None:
|
||||
attn = attn + attn_bias
|
||||
attn = attn.softmax(-1)
|
||||
return attn @ value
|
||||
|
||||
|
||||
class Attention(torch.nn.Module):
|
||||
|
||||
def __init__(self, q_dim, num_heads, head_dim, kv_dim=None, bias_q=False, bias_kv=False, bias_out=False):
|
||||
super().__init__()
|
||||
dim_inner = head_dim * num_heads
|
||||
kv_dim = kv_dim if kv_dim is not None else q_dim
|
||||
self.num_heads = num_heads
|
||||
self.head_dim = head_dim
|
||||
|
||||
self.to_q = torch.nn.Linear(q_dim, dim_inner, bias=bias_q)
|
||||
self.to_k = torch.nn.Linear(kv_dim, dim_inner, bias=bias_kv)
|
||||
self.to_v = torch.nn.Linear(kv_dim, dim_inner, bias=bias_kv)
|
||||
self.to_out = torch.nn.Linear(dim_inner, q_dim, bias=bias_out)
|
||||
|
||||
def interact_with_ipadapter(self, hidden_states, q, ip_k, ip_v, scale=1.0):
|
||||
batch_size = q.shape[0]
|
||||
ip_k = ip_k.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
|
||||
ip_v = ip_v.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
|
||||
ip_hidden_states = torch.nn.functional.scaled_dot_product_attention(q, ip_k, ip_v)
|
||||
hidden_states = hidden_states + scale * ip_hidden_states
|
||||
return hidden_states
|
||||
|
||||
def torch_forward(self, hidden_states, encoder_hidden_states=None, attn_mask=None, ipadapter_kwargs=None, qkv_preprocessor=None):
|
||||
if encoder_hidden_states is None:
|
||||
encoder_hidden_states = hidden_states
|
||||
|
||||
batch_size = encoder_hidden_states.shape[0]
|
||||
|
||||
q = self.to_q(hidden_states)
|
||||
k = self.to_k(encoder_hidden_states)
|
||||
v = self.to_v(encoder_hidden_states)
|
||||
|
||||
q = q.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
|
||||
k = k.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
|
||||
v = v.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
|
||||
|
||||
if qkv_preprocessor is not None:
|
||||
q, k, v = qkv_preprocessor(q, k, v)
|
||||
|
||||
hidden_states = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask)
|
||||
if ipadapter_kwargs is not None:
|
||||
hidden_states = self.interact_with_ipadapter(hidden_states, q, **ipadapter_kwargs)
|
||||
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, self.num_heads * self.head_dim)
|
||||
hidden_states = hidden_states.to(q.dtype)
|
||||
|
||||
hidden_states = self.to_out(hidden_states)
|
||||
|
||||
return hidden_states
|
||||
|
||||
def xformers_forward(self, hidden_states, encoder_hidden_states=None, attn_mask=None):
|
||||
if encoder_hidden_states is None:
|
||||
encoder_hidden_states = hidden_states
|
||||
|
||||
q = self.to_q(hidden_states)
|
||||
k = self.to_k(encoder_hidden_states)
|
||||
v = self.to_v(encoder_hidden_states)
|
||||
|
||||
q = rearrange(q, "b f (n d) -> (b n) f d", n=self.num_heads)
|
||||
k = rearrange(k, "b f (n d) -> (b n) f d", n=self.num_heads)
|
||||
v = rearrange(v, "b f (n d) -> (b n) f d", n=self.num_heads)
|
||||
|
||||
if attn_mask is not None:
|
||||
hidden_states = low_version_attention(q, k, v, attn_bias=attn_mask)
|
||||
else:
|
||||
import xformers.ops as xops
|
||||
hidden_states = xops.memory_efficient_attention(q, k, v)
|
||||
hidden_states = rearrange(hidden_states, "(b n) f d -> b f (n d)", n=self.num_heads)
|
||||
|
||||
hidden_states = hidden_states.to(q.dtype)
|
||||
hidden_states = self.to_out(hidden_states)
|
||||
|
||||
return hidden_states
|
||||
|
||||
def forward(self, hidden_states, encoder_hidden_states=None, attn_mask=None, ipadapter_kwargs=None, qkv_preprocessor=None):
|
||||
return self.torch_forward(hidden_states, encoder_hidden_states=encoder_hidden_states, attn_mask=attn_mask, ipadapter_kwargs=ipadapter_kwargs, qkv_preprocessor=qkv_preprocessor)
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
class CLIPEncoderLayer(torch.nn.Module):
|
||||
def __init__(self, embed_dim, intermediate_size, num_heads=12, head_dim=64, use_quick_gelu=True):
|
||||
super().__init__()
|
||||
self.attn = Attention(q_dim=embed_dim, num_heads=num_heads, head_dim=head_dim, bias_q=True, bias_kv=True, bias_out=True)
|
||||
self.layer_norm1 = torch.nn.LayerNorm(embed_dim)
|
||||
self.layer_norm2 = torch.nn.LayerNorm(embed_dim)
|
||||
self.fc1 = torch.nn.Linear(embed_dim, intermediate_size)
|
||||
self.fc2 = torch.nn.Linear(intermediate_size, embed_dim)
|
||||
|
||||
self.use_quick_gelu = use_quick_gelu
|
||||
|
||||
def quickGELU(self, x):
|
||||
return x * torch.sigmoid(1.702 * x)
|
||||
|
||||
def forward(self, hidden_states, attn_mask=None):
|
||||
residual = hidden_states
|
||||
|
||||
hidden_states = self.layer_norm1(hidden_states)
|
||||
hidden_states = self.attn(hidden_states, attn_mask=attn_mask)
|
||||
hidden_states = residual + hidden_states
|
||||
|
||||
residual = hidden_states
|
||||
hidden_states = self.layer_norm2(hidden_states)
|
||||
hidden_states = self.fc1(hidden_states)
|
||||
if self.use_quick_gelu:
|
||||
hidden_states = self.quickGELU(hidden_states)
|
||||
else:
|
||||
hidden_states = torch.nn.functional.gelu(hidden_states)
|
||||
hidden_states = self.fc2(hidden_states)
|
||||
hidden_states = residual + hidden_states
|
||||
|
||||
return hidden_states
|
||||
|
||||
|
||||
class SDTextEncoder(torch.nn.Module):
|
||||
def __init__(self, embed_dim=768, vocab_size=49408, max_position_embeddings=77, num_encoder_layers=12, encoder_intermediate_size=3072):
|
||||
super().__init__()
|
||||
|
||||
# token_embedding
|
||||
self.token_embedding = torch.nn.Embedding(vocab_size, embed_dim)
|
||||
|
||||
# position_embeds (This is a fixed tensor)
|
||||
self.position_embeds = torch.nn.Parameter(torch.zeros(1, max_position_embeddings, embed_dim))
|
||||
|
||||
# encoders
|
||||
self.encoders = torch.nn.ModuleList([CLIPEncoderLayer(embed_dim, encoder_intermediate_size) for _ in range(num_encoder_layers)])
|
||||
|
||||
# attn_mask
|
||||
self.attn_mask = self.attention_mask(max_position_embeddings)
|
||||
|
||||
# final_layer_norm
|
||||
self.final_layer_norm = torch.nn.LayerNorm(embed_dim)
|
||||
|
||||
def attention_mask(self, length):
|
||||
mask = torch.empty(length, length)
|
||||
mask.fill_(float("-inf"))
|
||||
mask.triu_(1)
|
||||
return mask
|
||||
|
||||
def forward(self, input_ids, clip_skip=1):
|
||||
embeds = self.token_embedding(input_ids) + self.position_embeds
|
||||
attn_mask = self.attn_mask.to(device=embeds.device, dtype=embeds.dtype)
|
||||
for encoder_id, encoder in enumerate(self.encoders):
|
||||
embeds = encoder(embeds, attn_mask=attn_mask)
|
||||
if encoder_id + clip_skip == len(self.encoders):
|
||||
break
|
||||
embeds = self.final_layer_norm(embeds)
|
||||
return embeds
|
||||
|
||||
@staticmethod
|
||||
def state_dict_converter():
|
||||
return SDTextEncoderStateDictConverter()
|
||||
|
||||
|
||||
class SDTextEncoderStateDictConverter:
|
||||
def __init__(self):
|
||||
pass
|
||||
|
||||
def from_diffusers(self, state_dict):
|
||||
rename_dict = {
|
||||
"text_model.embeddings.token_embedding.weight": "token_embedding.weight",
|
||||
"text_model.embeddings.position_embedding.weight": "position_embeds",
|
||||
"text_model.final_layer_norm.weight": "final_layer_norm.weight",
|
||||
"text_model.final_layer_norm.bias": "final_layer_norm.bias"
|
||||
}
|
||||
attn_rename_dict = {
|
||||
"self_attn.q_proj": "attn.to_q",
|
||||
"self_attn.k_proj": "attn.to_k",
|
||||
"self_attn.v_proj": "attn.to_v",
|
||||
"self_attn.out_proj": "attn.to_out",
|
||||
"layer_norm1": "layer_norm1",
|
||||
"layer_norm2": "layer_norm2",
|
||||
"mlp.fc1": "fc1",
|
||||
"mlp.fc2": "fc2",
|
||||
}
|
||||
state_dict_ = {}
|
||||
for name in state_dict:
|
||||
if name in rename_dict:
|
||||
param = state_dict[name]
|
||||
if name == "text_model.embeddings.position_embedding.weight":
|
||||
param = param.reshape((1, param.shape[0], param.shape[1]))
|
||||
state_dict_[rename_dict[name]] = param
|
||||
elif name.startswith("text_model.encoder.layers."):
|
||||
param = state_dict[name]
|
||||
names = name.split(".")
|
||||
layer_id, layer_type, tail = names[3], ".".join(names[4:-1]), names[-1]
|
||||
name_ = ".".join(["encoders", layer_id, attn_rename_dict[layer_type], tail])
|
||||
state_dict_[name_] = param
|
||||
return state_dict_
|
||||
|
||||
def from_civitai(self, state_dict):
|
||||
rename_dict = {
|
||||
"cond_stage_model.transformer.text_model.embeddings.token_embedding.weight": "token_embedding.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.layer_norm1.bias": "encoders.0.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.layer_norm1.weight": "encoders.0.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.layer_norm2.bias": "encoders.0.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.layer_norm2.weight": "encoders.0.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.mlp.fc1.bias": "encoders.0.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.mlp.fc1.weight": "encoders.0.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.mlp.fc2.bias": "encoders.0.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.mlp.fc2.weight": "encoders.0.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.k_proj.bias": "encoders.0.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.k_proj.weight": "encoders.0.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.out_proj.bias": "encoders.0.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.out_proj.weight": "encoders.0.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.q_proj.bias": "encoders.0.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.q_proj.weight": "encoders.0.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.v_proj.bias": "encoders.0.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.v_proj.weight": "encoders.0.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.layer_norm1.bias": "encoders.1.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.layer_norm1.weight": "encoders.1.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.layer_norm2.bias": "encoders.1.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.layer_norm2.weight": "encoders.1.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.mlp.fc1.bias": "encoders.1.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.mlp.fc1.weight": "encoders.1.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.mlp.fc2.bias": "encoders.1.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.mlp.fc2.weight": "encoders.1.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.k_proj.bias": "encoders.1.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.k_proj.weight": "encoders.1.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.out_proj.bias": "encoders.1.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.out_proj.weight": "encoders.1.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.q_proj.bias": "encoders.1.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.q_proj.weight": "encoders.1.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.v_proj.bias": "encoders.1.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.v_proj.weight": "encoders.1.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.layer_norm1.bias": "encoders.10.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.layer_norm1.weight": "encoders.10.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.layer_norm2.bias": "encoders.10.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.layer_norm2.weight": "encoders.10.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.mlp.fc1.bias": "encoders.10.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.mlp.fc1.weight": "encoders.10.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.mlp.fc2.bias": "encoders.10.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.mlp.fc2.weight": "encoders.10.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.k_proj.bias": "encoders.10.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.k_proj.weight": "encoders.10.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.out_proj.bias": "encoders.10.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.out_proj.weight": "encoders.10.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.q_proj.bias": "encoders.10.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.q_proj.weight": "encoders.10.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.v_proj.bias": "encoders.10.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.v_proj.weight": "encoders.10.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.layer_norm1.bias": "encoders.11.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.layer_norm1.weight": "encoders.11.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.layer_norm2.bias": "encoders.11.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.layer_norm2.weight": "encoders.11.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.mlp.fc1.bias": "encoders.11.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.mlp.fc1.weight": "encoders.11.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.mlp.fc2.bias": "encoders.11.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.mlp.fc2.weight": "encoders.11.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.k_proj.bias": "encoders.11.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.k_proj.weight": "encoders.11.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.out_proj.bias": "encoders.11.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.out_proj.weight": "encoders.11.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.q_proj.bias": "encoders.11.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.q_proj.weight": "encoders.11.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.v_proj.bias": "encoders.11.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.v_proj.weight": "encoders.11.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.layer_norm1.bias": "encoders.2.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.layer_norm1.weight": "encoders.2.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.layer_norm2.bias": "encoders.2.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.layer_norm2.weight": "encoders.2.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.mlp.fc1.bias": "encoders.2.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.mlp.fc1.weight": "encoders.2.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.mlp.fc2.bias": "encoders.2.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.mlp.fc2.weight": "encoders.2.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.k_proj.bias": "encoders.2.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.k_proj.weight": "encoders.2.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.out_proj.bias": "encoders.2.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.out_proj.weight": "encoders.2.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.q_proj.bias": "encoders.2.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.q_proj.weight": "encoders.2.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.v_proj.bias": "encoders.2.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.v_proj.weight": "encoders.2.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.layer_norm1.bias": "encoders.3.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.layer_norm1.weight": "encoders.3.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.layer_norm2.bias": "encoders.3.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.layer_norm2.weight": "encoders.3.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.mlp.fc1.bias": "encoders.3.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.mlp.fc1.weight": "encoders.3.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.mlp.fc2.bias": "encoders.3.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.mlp.fc2.weight": "encoders.3.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.k_proj.bias": "encoders.3.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.k_proj.weight": "encoders.3.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.out_proj.bias": "encoders.3.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.out_proj.weight": "encoders.3.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.q_proj.bias": "encoders.3.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.q_proj.weight": "encoders.3.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.v_proj.bias": "encoders.3.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.v_proj.weight": "encoders.3.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.layer_norm1.bias": "encoders.4.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.layer_norm1.weight": "encoders.4.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.layer_norm2.bias": "encoders.4.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.layer_norm2.weight": "encoders.4.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.mlp.fc1.bias": "encoders.4.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.mlp.fc1.weight": "encoders.4.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.mlp.fc2.bias": "encoders.4.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.mlp.fc2.weight": "encoders.4.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.k_proj.bias": "encoders.4.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.k_proj.weight": "encoders.4.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.out_proj.bias": "encoders.4.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.out_proj.weight": "encoders.4.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.q_proj.bias": "encoders.4.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.q_proj.weight": "encoders.4.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.v_proj.bias": "encoders.4.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.v_proj.weight": "encoders.4.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.layer_norm1.bias": "encoders.5.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.layer_norm1.weight": "encoders.5.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.layer_norm2.bias": "encoders.5.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.layer_norm2.weight": "encoders.5.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.mlp.fc1.bias": "encoders.5.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.mlp.fc1.weight": "encoders.5.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.mlp.fc2.bias": "encoders.5.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.mlp.fc2.weight": "encoders.5.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.k_proj.bias": "encoders.5.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.k_proj.weight": "encoders.5.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.out_proj.bias": "encoders.5.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.out_proj.weight": "encoders.5.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.q_proj.bias": "encoders.5.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.q_proj.weight": "encoders.5.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.v_proj.bias": "encoders.5.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.v_proj.weight": "encoders.5.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.layer_norm1.bias": "encoders.6.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.layer_norm1.weight": "encoders.6.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.layer_norm2.bias": "encoders.6.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.layer_norm2.weight": "encoders.6.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.mlp.fc1.bias": "encoders.6.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.mlp.fc1.weight": "encoders.6.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.mlp.fc2.bias": "encoders.6.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.mlp.fc2.weight": "encoders.6.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.k_proj.bias": "encoders.6.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.k_proj.weight": "encoders.6.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.out_proj.bias": "encoders.6.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.out_proj.weight": "encoders.6.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.q_proj.bias": "encoders.6.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.q_proj.weight": "encoders.6.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.v_proj.bias": "encoders.6.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.v_proj.weight": "encoders.6.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.layer_norm1.bias": "encoders.7.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.layer_norm1.weight": "encoders.7.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.layer_norm2.bias": "encoders.7.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.layer_norm2.weight": "encoders.7.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.mlp.fc1.bias": "encoders.7.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.mlp.fc1.weight": "encoders.7.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.mlp.fc2.bias": "encoders.7.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.mlp.fc2.weight": "encoders.7.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.k_proj.bias": "encoders.7.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.k_proj.weight": "encoders.7.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.out_proj.bias": "encoders.7.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.out_proj.weight": "encoders.7.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.q_proj.bias": "encoders.7.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.q_proj.weight": "encoders.7.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.v_proj.bias": "encoders.7.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.v_proj.weight": "encoders.7.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.layer_norm1.bias": "encoders.8.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.layer_norm1.weight": "encoders.8.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.layer_norm2.bias": "encoders.8.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.layer_norm2.weight": "encoders.8.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.mlp.fc1.bias": "encoders.8.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.mlp.fc1.weight": "encoders.8.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.mlp.fc2.bias": "encoders.8.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.mlp.fc2.weight": "encoders.8.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.k_proj.bias": "encoders.8.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.k_proj.weight": "encoders.8.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.out_proj.bias": "encoders.8.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.out_proj.weight": "encoders.8.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.q_proj.bias": "encoders.8.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.q_proj.weight": "encoders.8.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.v_proj.bias": "encoders.8.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.v_proj.weight": "encoders.8.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.layer_norm1.bias": "encoders.9.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.layer_norm1.weight": "encoders.9.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.layer_norm2.bias": "encoders.9.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.layer_norm2.weight": "encoders.9.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.mlp.fc1.bias": "encoders.9.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.mlp.fc1.weight": "encoders.9.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.mlp.fc2.bias": "encoders.9.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.mlp.fc2.weight": "encoders.9.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.k_proj.bias": "encoders.9.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.k_proj.weight": "encoders.9.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.out_proj.bias": "encoders.9.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.out_proj.weight": "encoders.9.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.q_proj.bias": "encoders.9.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.q_proj.weight": "encoders.9.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.v_proj.bias": "encoders.9.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.v_proj.weight": "encoders.9.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.final_layer_norm.bias": "final_layer_norm.bias",
|
||||
"cond_stage_model.transformer.text_model.final_layer_norm.weight": "final_layer_norm.weight",
|
||||
"cond_stage_model.transformer.text_model.embeddings.position_embedding.weight": "position_embeds"
|
||||
}
|
||||
state_dict_ = {}
|
||||
for name in state_dict:
|
||||
if name in rename_dict:
|
||||
param = state_dict[name]
|
||||
if name == "cond_stage_model.transformer.text_model.embeddings.position_embedding.weight":
|
||||
param = param.reshape((1, param.shape[0], param.shape[1]))
|
||||
state_dict_[rename_dict[name]] = param
|
||||
return state_dict_
|
||||
|
||||
|
||||
|
||||
class LoRALayerBlock(torch.nn.Module):
|
||||
@@ -58,13 +468,80 @@ class LoRAEmbedder(torch.nn.Module):
|
||||
"type": suffix,
|
||||
})
|
||||
return lora_patterns
|
||||
|
||||
def get_lora_param_pair(self, lora, name, dim, device, dtype):
|
||||
key_A = name + ".lora_A.default.weight"
|
||||
key_B = name + ".lora_B.default.weight"
|
||||
if key_A in lora and key_B in lora:
|
||||
return lora[key_A], lora[key_B]
|
||||
|
||||
if "to_qkv" in name:
|
||||
base_name = name.replace("to_qkv", "")
|
||||
suffixes = ["to_q", "to_k", "to_v"]
|
||||
|
||||
found_As = []
|
||||
found_Bs = []
|
||||
|
||||
all_found = True
|
||||
for suffix in suffixes:
|
||||
sub_name = base_name + suffix
|
||||
k_A = sub_name + ".lora_A.default.weight"
|
||||
k_B = sub_name + ".lora_B.default.weight"
|
||||
|
||||
if k_A in lora and k_B in lora:
|
||||
found_As.append(lora[k_A])
|
||||
found_Bs.append(lora[k_B])
|
||||
else:
|
||||
all_found = False
|
||||
break
|
||||
if all_found:
|
||||
pass
|
||||
|
||||
rank = 16
|
||||
for k, v in lora.items():
|
||||
if "lora_A" in k:
|
||||
rank = v.shape[0]
|
||||
device = v.device
|
||||
dtype = v.dtype
|
||||
break
|
||||
|
||||
lora_A = torch.zeros((rank, dim[0]), device=device, dtype=dtype)
|
||||
lora_B = torch.zeros((dim[1], rank), device=device, dtype=dtype)
|
||||
|
||||
return lora_A, lora_B
|
||||
|
||||
def forward(self, lora):
|
||||
lora_emb = []
|
||||
device = None
|
||||
dtype = None
|
||||
for v in lora.values():
|
||||
device = v.device
|
||||
dtype = v.dtype
|
||||
break
|
||||
|
||||
for lora_pattern in self.lora_patterns:
|
||||
name, layer_type = lora_pattern["name"], lora_pattern["type"]
|
||||
lora_A = lora[name + ".lora_A.default.weight"]
|
||||
lora_B = lora[name + ".lora_B.default.weight"]
|
||||
dim = lora_pattern["dim"]
|
||||
|
||||
lora_A, lora_B = self.get_lora_param_pair(lora, name, dim, device, dtype)
|
||||
|
||||
if "to_qkv" in name and (lora_A is None or (torch.equal(lora_A, torch.zeros_like(lora_A)))):
|
||||
base_name = name.replace("to_qkv", "")
|
||||
try:
|
||||
q_name = base_name + "to_q"
|
||||
k_name = base_name + "to_k"
|
||||
v_name = base_name + "to_v"
|
||||
|
||||
real_A = lora[q_name + ".lora_A.default.weight"]
|
||||
B_q = lora[q_name + ".lora_B.default.weight"]
|
||||
B_k = lora[k_name + ".lora_B.default.weight"]
|
||||
B_v = lora[v_name + ".lora_B.default.weight"]
|
||||
real_B = torch.cat([B_q, B_k, B_v], dim=0)
|
||||
|
||||
lora_A, lora_B = real_A, real_B
|
||||
except KeyError:
|
||||
pass
|
||||
|
||||
lora_out = self.model_dict[name.replace(".", "___")](lora_A, lora_B)
|
||||
lora_out = self.proj_dict[layer_type.replace(".", "___")](lora_out)
|
||||
lora_emb.append(lora_out)
|
||||
|
||||
60
diffsynth/models/flux_lora_patcher.py
Normal file
60
diffsynth/models/flux_lora_patcher.py
Normal file
@@ -0,0 +1,60 @@
|
||||
import torch
|
||||
|
||||
class LoraMerger(torch.nn.Module):
|
||||
def __init__(self, dim):
|
||||
super().__init__()
|
||||
self.weight_base = torch.nn.Parameter(torch.randn((dim,)))
|
||||
self.weight_lora = torch.nn.Parameter(torch.randn((dim,)))
|
||||
self.weight_cross = torch.nn.Parameter(torch.randn((dim,)))
|
||||
self.weight_out = torch.nn.Parameter(torch.ones((dim,)))
|
||||
self.bias = torch.nn.Parameter(torch.randn((dim,)))
|
||||
self.activation = torch.nn.Sigmoid()
|
||||
self.norm_base = torch.nn.LayerNorm(dim, eps=1e-5)
|
||||
self.norm_lora = torch.nn.LayerNorm(dim, eps=1e-5)
|
||||
|
||||
def forward(self, base_output, lora_outputs):
|
||||
norm_base_output = self.norm_base(base_output)
|
||||
norm_lora_outputs = self.norm_lora(lora_outputs)
|
||||
gate = self.activation(
|
||||
norm_base_output * self.weight_base \
|
||||
+ norm_lora_outputs * self.weight_lora \
|
||||
+ norm_base_output * norm_lora_outputs * self.weight_cross + self.bias
|
||||
)
|
||||
output = base_output + (self.weight_out * gate * lora_outputs).sum(dim=0)
|
||||
return output
|
||||
|
||||
class FluxLoraPatcher(torch.nn.Module):
|
||||
def __init__(self, lora_patterns=None):
|
||||
super().__init__()
|
||||
if lora_patterns is None:
|
||||
lora_patterns = self.default_lora_patterns()
|
||||
model_dict = {}
|
||||
for lora_pattern in lora_patterns:
|
||||
name, dim = lora_pattern["name"], lora_pattern["dim"]
|
||||
model_dict[name.replace(".", "___")] = LoraMerger(dim)
|
||||
self.model_dict = torch.nn.ModuleDict(model_dict)
|
||||
|
||||
def default_lora_patterns(self):
|
||||
lora_patterns = []
|
||||
lora_dict = {
|
||||
"attn.a_to_qkv": 9216, "attn.a_to_out": 3072, "ff_a.0": 12288, "ff_a.2": 3072, "norm1_a.linear": 18432,
|
||||
"attn.b_to_qkv": 9216, "attn.b_to_out": 3072, "ff_b.0": 12288, "ff_b.2": 3072, "norm1_b.linear": 18432,
|
||||
}
|
||||
for i in range(19):
|
||||
for suffix in lora_dict:
|
||||
lora_patterns.append({
|
||||
"name": f"blocks.{i}.{suffix}",
|
||||
"dim": lora_dict[suffix]
|
||||
})
|
||||
lora_dict = {"to_qkv_mlp": 21504, "proj_out": 3072, "norm.linear": 9216}
|
||||
for i in range(38):
|
||||
for suffix in lora_dict:
|
||||
lora_patterns.append({
|
||||
"name": f"single_blocks.{i}.{suffix}",
|
||||
"dim": lora_dict[suffix]
|
||||
})
|
||||
return lora_patterns
|
||||
|
||||
def forward(self, base_output, lora_outputs, name):
|
||||
return self.model_dict[name.replace(".", "___")](base_output, lora_outputs)
|
||||
|
||||
412
diffsynth/models/sd_text_encoder.py
Normal file
412
diffsynth/models/sd_text_encoder.py
Normal file
@@ -0,0 +1,412 @@
|
||||
import torch
|
||||
from .attention import Attention
|
||||
from einops import rearrange
|
||||
|
||||
|
||||
def low_version_attention(query, key, value, attn_bias=None):
|
||||
scale = 1 / query.shape[-1] ** 0.5
|
||||
query = query * scale
|
||||
attn = torch.matmul(query, key.transpose(-2, -1))
|
||||
if attn_bias is not None:
|
||||
attn = attn + attn_bias
|
||||
attn = attn.softmax(-1)
|
||||
return attn @ value
|
||||
|
||||
|
||||
class Attention(torch.nn.Module):
|
||||
|
||||
def __init__(self, q_dim, num_heads, head_dim, kv_dim=None, bias_q=False, bias_kv=False, bias_out=False):
|
||||
super().__init__()
|
||||
dim_inner = head_dim * num_heads
|
||||
kv_dim = kv_dim if kv_dim is not None else q_dim
|
||||
self.num_heads = num_heads
|
||||
self.head_dim = head_dim
|
||||
|
||||
self.to_q = torch.nn.Linear(q_dim, dim_inner, bias=bias_q)
|
||||
self.to_k = torch.nn.Linear(kv_dim, dim_inner, bias=bias_kv)
|
||||
self.to_v = torch.nn.Linear(kv_dim, dim_inner, bias=bias_kv)
|
||||
self.to_out = torch.nn.Linear(dim_inner, q_dim, bias=bias_out)
|
||||
|
||||
def interact_with_ipadapter(self, hidden_states, q, ip_k, ip_v, scale=1.0):
|
||||
batch_size = q.shape[0]
|
||||
ip_k = ip_k.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
|
||||
ip_v = ip_v.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
|
||||
ip_hidden_states = torch.nn.functional.scaled_dot_product_attention(q, ip_k, ip_v)
|
||||
hidden_states = hidden_states + scale * ip_hidden_states
|
||||
return hidden_states
|
||||
|
||||
def torch_forward(self, hidden_states, encoder_hidden_states=None, attn_mask=None, ipadapter_kwargs=None, qkv_preprocessor=None):
|
||||
if encoder_hidden_states is None:
|
||||
encoder_hidden_states = hidden_states
|
||||
|
||||
batch_size = encoder_hidden_states.shape[0]
|
||||
|
||||
q = self.to_q(hidden_states)
|
||||
k = self.to_k(encoder_hidden_states)
|
||||
v = self.to_v(encoder_hidden_states)
|
||||
|
||||
q = q.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
|
||||
k = k.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
|
||||
v = v.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
|
||||
|
||||
if qkv_preprocessor is not None:
|
||||
q, k, v = qkv_preprocessor(q, k, v)
|
||||
|
||||
hidden_states = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask)
|
||||
if ipadapter_kwargs is not None:
|
||||
hidden_states = self.interact_with_ipadapter(hidden_states, q, **ipadapter_kwargs)
|
||||
hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, self.num_heads * self.head_dim)
|
||||
hidden_states = hidden_states.to(q.dtype)
|
||||
|
||||
hidden_states = self.to_out(hidden_states)
|
||||
|
||||
return hidden_states
|
||||
|
||||
def xformers_forward(self, hidden_states, encoder_hidden_states=None, attn_mask=None):
|
||||
if encoder_hidden_states is None:
|
||||
encoder_hidden_states = hidden_states
|
||||
|
||||
q = self.to_q(hidden_states)
|
||||
k = self.to_k(encoder_hidden_states)
|
||||
v = self.to_v(encoder_hidden_states)
|
||||
|
||||
q = rearrange(q, "b f (n d) -> (b n) f d", n=self.num_heads)
|
||||
k = rearrange(k, "b f (n d) -> (b n) f d", n=self.num_heads)
|
||||
v = rearrange(v, "b f (n d) -> (b n) f d", n=self.num_heads)
|
||||
|
||||
if attn_mask is not None:
|
||||
hidden_states = low_version_attention(q, k, v, attn_bias=attn_mask)
|
||||
else:
|
||||
import xformers.ops as xops
|
||||
hidden_states = xops.memory_efficient_attention(q, k, v)
|
||||
hidden_states = rearrange(hidden_states, "(b n) f d -> b f (n d)", n=self.num_heads)
|
||||
|
||||
hidden_states = hidden_states.to(q.dtype)
|
||||
hidden_states = self.to_out(hidden_states)
|
||||
|
||||
return hidden_states
|
||||
|
||||
def forward(self, hidden_states, encoder_hidden_states=None, attn_mask=None, ipadapter_kwargs=None, qkv_preprocessor=None):
|
||||
return self.torch_forward(hidden_states, encoder_hidden_states=encoder_hidden_states, attn_mask=attn_mask, ipadapter_kwargs=ipadapter_kwargs, qkv_preprocessor=qkv_preprocessor)
|
||||
|
||||
|
||||
|
||||
|
||||
|
||||
class CLIPEncoderLayer(torch.nn.Module):
|
||||
def __init__(self, embed_dim, intermediate_size, num_heads=12, head_dim=64, use_quick_gelu=True):
|
||||
super().__init__()
|
||||
self.attn = Attention(q_dim=embed_dim, num_heads=num_heads, head_dim=head_dim, bias_q=True, bias_kv=True, bias_out=True)
|
||||
self.layer_norm1 = torch.nn.LayerNorm(embed_dim)
|
||||
self.layer_norm2 = torch.nn.LayerNorm(embed_dim)
|
||||
self.fc1 = torch.nn.Linear(embed_dim, intermediate_size)
|
||||
self.fc2 = torch.nn.Linear(intermediate_size, embed_dim)
|
||||
|
||||
self.use_quick_gelu = use_quick_gelu
|
||||
|
||||
def quickGELU(self, x):
|
||||
return x * torch.sigmoid(1.702 * x)
|
||||
|
||||
def forward(self, hidden_states, attn_mask=None):
|
||||
residual = hidden_states
|
||||
|
||||
hidden_states = self.layer_norm1(hidden_states)
|
||||
hidden_states = self.attn(hidden_states, attn_mask=attn_mask)
|
||||
hidden_states = residual + hidden_states
|
||||
|
||||
residual = hidden_states
|
||||
hidden_states = self.layer_norm2(hidden_states)
|
||||
hidden_states = self.fc1(hidden_states)
|
||||
if self.use_quick_gelu:
|
||||
hidden_states = self.quickGELU(hidden_states)
|
||||
else:
|
||||
hidden_states = torch.nn.functional.gelu(hidden_states)
|
||||
hidden_states = self.fc2(hidden_states)
|
||||
hidden_states = residual + hidden_states
|
||||
|
||||
return hidden_states
|
||||
|
||||
|
||||
class SDTextEncoder(torch.nn.Module):
|
||||
def __init__(self, embed_dim=768, vocab_size=49408, max_position_embeddings=77, num_encoder_layers=12, encoder_intermediate_size=3072):
|
||||
super().__init__()
|
||||
|
||||
# token_embedding
|
||||
self.token_embedding = torch.nn.Embedding(vocab_size, embed_dim)
|
||||
|
||||
# position_embeds (This is a fixed tensor)
|
||||
self.position_embeds = torch.nn.Parameter(torch.zeros(1, max_position_embeddings, embed_dim))
|
||||
|
||||
# encoders
|
||||
self.encoders = torch.nn.ModuleList([CLIPEncoderLayer(embed_dim, encoder_intermediate_size) for _ in range(num_encoder_layers)])
|
||||
|
||||
# attn_mask
|
||||
self.attn_mask = self.attention_mask(max_position_embeddings)
|
||||
|
||||
# final_layer_norm
|
||||
self.final_layer_norm = torch.nn.LayerNorm(embed_dim)
|
||||
|
||||
def attention_mask(self, length):
|
||||
mask = torch.empty(length, length)
|
||||
mask.fill_(float("-inf"))
|
||||
mask.triu_(1)
|
||||
return mask
|
||||
|
||||
def forward(self, input_ids, clip_skip=1):
|
||||
embeds = self.token_embedding(input_ids) + self.position_embeds
|
||||
attn_mask = self.attn_mask.to(device=embeds.device, dtype=embeds.dtype)
|
||||
for encoder_id, encoder in enumerate(self.encoders):
|
||||
embeds = encoder(embeds, attn_mask=attn_mask)
|
||||
if encoder_id + clip_skip == len(self.encoders):
|
||||
break
|
||||
embeds = self.final_layer_norm(embeds)
|
||||
return embeds
|
||||
|
||||
@staticmethod
|
||||
def state_dict_converter():
|
||||
return SDTextEncoderStateDictConverter()
|
||||
|
||||
|
||||
class SDTextEncoderStateDictConverter:
|
||||
def __init__(self):
|
||||
pass
|
||||
|
||||
def from_diffusers(self, state_dict):
|
||||
rename_dict = {
|
||||
"text_model.embeddings.token_embedding.weight": "token_embedding.weight",
|
||||
"text_model.embeddings.position_embedding.weight": "position_embeds",
|
||||
"text_model.final_layer_norm.weight": "final_layer_norm.weight",
|
||||
"text_model.final_layer_norm.bias": "final_layer_norm.bias"
|
||||
}
|
||||
attn_rename_dict = {
|
||||
"self_attn.q_proj": "attn.to_q",
|
||||
"self_attn.k_proj": "attn.to_k",
|
||||
"self_attn.v_proj": "attn.to_v",
|
||||
"self_attn.out_proj": "attn.to_out",
|
||||
"layer_norm1": "layer_norm1",
|
||||
"layer_norm2": "layer_norm2",
|
||||
"mlp.fc1": "fc1",
|
||||
"mlp.fc2": "fc2",
|
||||
}
|
||||
state_dict_ = {}
|
||||
for name in state_dict:
|
||||
if name in rename_dict:
|
||||
param = state_dict[name]
|
||||
if name == "text_model.embeddings.position_embedding.weight":
|
||||
param = param.reshape((1, param.shape[0], param.shape[1]))
|
||||
state_dict_[rename_dict[name]] = param
|
||||
elif name.startswith("text_model.encoder.layers."):
|
||||
param = state_dict[name]
|
||||
names = name.split(".")
|
||||
layer_id, layer_type, tail = names[3], ".".join(names[4:-1]), names[-1]
|
||||
name_ = ".".join(["encoders", layer_id, attn_rename_dict[layer_type], tail])
|
||||
state_dict_[name_] = param
|
||||
return state_dict_
|
||||
|
||||
def from_civitai(self, state_dict):
|
||||
rename_dict = {
|
||||
"cond_stage_model.transformer.text_model.embeddings.token_embedding.weight": "token_embedding.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.layer_norm1.bias": "encoders.0.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.layer_norm1.weight": "encoders.0.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.layer_norm2.bias": "encoders.0.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.layer_norm2.weight": "encoders.0.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.mlp.fc1.bias": "encoders.0.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.mlp.fc1.weight": "encoders.0.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.mlp.fc2.bias": "encoders.0.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.mlp.fc2.weight": "encoders.0.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.k_proj.bias": "encoders.0.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.k_proj.weight": "encoders.0.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.out_proj.bias": "encoders.0.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.out_proj.weight": "encoders.0.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.q_proj.bias": "encoders.0.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.q_proj.weight": "encoders.0.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.v_proj.bias": "encoders.0.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.0.self_attn.v_proj.weight": "encoders.0.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.layer_norm1.bias": "encoders.1.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.layer_norm1.weight": "encoders.1.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.layer_norm2.bias": "encoders.1.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.layer_norm2.weight": "encoders.1.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.mlp.fc1.bias": "encoders.1.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.mlp.fc1.weight": "encoders.1.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.mlp.fc2.bias": "encoders.1.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.mlp.fc2.weight": "encoders.1.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.k_proj.bias": "encoders.1.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.k_proj.weight": "encoders.1.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.out_proj.bias": "encoders.1.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.out_proj.weight": "encoders.1.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.q_proj.bias": "encoders.1.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.q_proj.weight": "encoders.1.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.v_proj.bias": "encoders.1.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.1.self_attn.v_proj.weight": "encoders.1.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.layer_norm1.bias": "encoders.10.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.layer_norm1.weight": "encoders.10.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.layer_norm2.bias": "encoders.10.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.layer_norm2.weight": "encoders.10.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.mlp.fc1.bias": "encoders.10.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.mlp.fc1.weight": "encoders.10.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.mlp.fc2.bias": "encoders.10.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.mlp.fc2.weight": "encoders.10.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.k_proj.bias": "encoders.10.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.k_proj.weight": "encoders.10.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.out_proj.bias": "encoders.10.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.out_proj.weight": "encoders.10.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.q_proj.bias": "encoders.10.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.q_proj.weight": "encoders.10.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.v_proj.bias": "encoders.10.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.10.self_attn.v_proj.weight": "encoders.10.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.layer_norm1.bias": "encoders.11.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.layer_norm1.weight": "encoders.11.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.layer_norm2.bias": "encoders.11.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.layer_norm2.weight": "encoders.11.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.mlp.fc1.bias": "encoders.11.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.mlp.fc1.weight": "encoders.11.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.mlp.fc2.bias": "encoders.11.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.mlp.fc2.weight": "encoders.11.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.k_proj.bias": "encoders.11.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.k_proj.weight": "encoders.11.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.out_proj.bias": "encoders.11.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.out_proj.weight": "encoders.11.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.q_proj.bias": "encoders.11.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.q_proj.weight": "encoders.11.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.v_proj.bias": "encoders.11.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.11.self_attn.v_proj.weight": "encoders.11.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.layer_norm1.bias": "encoders.2.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.layer_norm1.weight": "encoders.2.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.layer_norm2.bias": "encoders.2.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.layer_norm2.weight": "encoders.2.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.mlp.fc1.bias": "encoders.2.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.mlp.fc1.weight": "encoders.2.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.mlp.fc2.bias": "encoders.2.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.mlp.fc2.weight": "encoders.2.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.k_proj.bias": "encoders.2.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.k_proj.weight": "encoders.2.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.out_proj.bias": "encoders.2.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.out_proj.weight": "encoders.2.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.q_proj.bias": "encoders.2.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.q_proj.weight": "encoders.2.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.v_proj.bias": "encoders.2.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.2.self_attn.v_proj.weight": "encoders.2.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.layer_norm1.bias": "encoders.3.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.layer_norm1.weight": "encoders.3.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.layer_norm2.bias": "encoders.3.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.layer_norm2.weight": "encoders.3.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.mlp.fc1.bias": "encoders.3.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.mlp.fc1.weight": "encoders.3.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.mlp.fc2.bias": "encoders.3.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.mlp.fc2.weight": "encoders.3.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.k_proj.bias": "encoders.3.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.k_proj.weight": "encoders.3.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.out_proj.bias": "encoders.3.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.out_proj.weight": "encoders.3.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.q_proj.bias": "encoders.3.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.q_proj.weight": "encoders.3.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.v_proj.bias": "encoders.3.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.3.self_attn.v_proj.weight": "encoders.3.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.layer_norm1.bias": "encoders.4.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.layer_norm1.weight": "encoders.4.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.layer_norm2.bias": "encoders.4.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.layer_norm2.weight": "encoders.4.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.mlp.fc1.bias": "encoders.4.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.mlp.fc1.weight": "encoders.4.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.mlp.fc2.bias": "encoders.4.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.mlp.fc2.weight": "encoders.4.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.k_proj.bias": "encoders.4.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.k_proj.weight": "encoders.4.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.out_proj.bias": "encoders.4.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.out_proj.weight": "encoders.4.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.q_proj.bias": "encoders.4.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.q_proj.weight": "encoders.4.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.v_proj.bias": "encoders.4.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.4.self_attn.v_proj.weight": "encoders.4.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.layer_norm1.bias": "encoders.5.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.layer_norm1.weight": "encoders.5.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.layer_norm2.bias": "encoders.5.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.layer_norm2.weight": "encoders.5.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.mlp.fc1.bias": "encoders.5.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.mlp.fc1.weight": "encoders.5.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.mlp.fc2.bias": "encoders.5.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.mlp.fc2.weight": "encoders.5.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.k_proj.bias": "encoders.5.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.k_proj.weight": "encoders.5.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.out_proj.bias": "encoders.5.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.out_proj.weight": "encoders.5.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.q_proj.bias": "encoders.5.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.q_proj.weight": "encoders.5.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.v_proj.bias": "encoders.5.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.5.self_attn.v_proj.weight": "encoders.5.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.layer_norm1.bias": "encoders.6.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.layer_norm1.weight": "encoders.6.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.layer_norm2.bias": "encoders.6.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.layer_norm2.weight": "encoders.6.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.mlp.fc1.bias": "encoders.6.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.mlp.fc1.weight": "encoders.6.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.mlp.fc2.bias": "encoders.6.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.mlp.fc2.weight": "encoders.6.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.k_proj.bias": "encoders.6.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.k_proj.weight": "encoders.6.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.out_proj.bias": "encoders.6.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.out_proj.weight": "encoders.6.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.q_proj.bias": "encoders.6.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.q_proj.weight": "encoders.6.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.v_proj.bias": "encoders.6.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.6.self_attn.v_proj.weight": "encoders.6.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.layer_norm1.bias": "encoders.7.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.layer_norm1.weight": "encoders.7.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.layer_norm2.bias": "encoders.7.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.layer_norm2.weight": "encoders.7.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.mlp.fc1.bias": "encoders.7.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.mlp.fc1.weight": "encoders.7.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.mlp.fc2.bias": "encoders.7.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.mlp.fc2.weight": "encoders.7.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.k_proj.bias": "encoders.7.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.k_proj.weight": "encoders.7.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.out_proj.bias": "encoders.7.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.out_proj.weight": "encoders.7.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.q_proj.bias": "encoders.7.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.q_proj.weight": "encoders.7.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.v_proj.bias": "encoders.7.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.7.self_attn.v_proj.weight": "encoders.7.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.layer_norm1.bias": "encoders.8.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.layer_norm1.weight": "encoders.8.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.layer_norm2.bias": "encoders.8.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.layer_norm2.weight": "encoders.8.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.mlp.fc1.bias": "encoders.8.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.mlp.fc1.weight": "encoders.8.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.mlp.fc2.bias": "encoders.8.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.mlp.fc2.weight": "encoders.8.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.k_proj.bias": "encoders.8.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.k_proj.weight": "encoders.8.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.out_proj.bias": "encoders.8.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.out_proj.weight": "encoders.8.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.q_proj.bias": "encoders.8.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.q_proj.weight": "encoders.8.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.v_proj.bias": "encoders.8.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.8.self_attn.v_proj.weight": "encoders.8.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.layer_norm1.bias": "encoders.9.layer_norm1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.layer_norm1.weight": "encoders.9.layer_norm1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.layer_norm2.bias": "encoders.9.layer_norm2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.layer_norm2.weight": "encoders.9.layer_norm2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.mlp.fc1.bias": "encoders.9.fc1.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.mlp.fc1.weight": "encoders.9.fc1.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.mlp.fc2.bias": "encoders.9.fc2.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.mlp.fc2.weight": "encoders.9.fc2.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.k_proj.bias": "encoders.9.attn.to_k.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.k_proj.weight": "encoders.9.attn.to_k.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.out_proj.bias": "encoders.9.attn.to_out.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.out_proj.weight": "encoders.9.attn.to_out.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.q_proj.bias": "encoders.9.attn.to_q.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.q_proj.weight": "encoders.9.attn.to_q.weight",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.v_proj.bias": "encoders.9.attn.to_v.bias",
|
||||
"cond_stage_model.transformer.text_model.encoder.layers.9.self_attn.v_proj.weight": "encoders.9.attn.to_v.weight",
|
||||
"cond_stage_model.transformer.text_model.final_layer_norm.bias": "final_layer_norm.bias",
|
||||
"cond_stage_model.transformer.text_model.final_layer_norm.weight": "final_layer_norm.weight",
|
||||
"cond_stage_model.transformer.text_model.embeddings.position_embedding.weight": "position_embeds"
|
||||
}
|
||||
state_dict_ = {}
|
||||
for name in state_dict:
|
||||
if name in rename_dict:
|
||||
param = state_dict[name]
|
||||
if name == "cond_stage_model.transformer.text_model.embeddings.position_embedding.weight":
|
||||
param = param.reshape((1, param.shape[0], param.shape[1]))
|
||||
state_dict_[rename_dict[name]] = param
|
||||
return state_dict_
|
||||
Reference in New Issue
Block a user