Upload folder using huggingface_hub
Browse files- .gitattributes +1 -0
- llm_adapter/config.json +12 -0
- llm_adapter/diffusion_pytorch_model.safetensors +3 -0
- llm_adapter/modeling_llm_adapter.py +215 -0
- model_index.json +35 -0
- pipeline.py +371 -0
- scheduler/scheduler_config.json +6 -0
- t5_tokenizer/tokenizer.json +0 -0
- t5_tokenizer/tokenizer_config.json +113 -0
- text_encoder/config.json +22 -0
- text_encoder/model.safetensors +3 -0
- tokenizer/chat_template.jinja +89 -0
- tokenizer/tokenizer.json +3 -0
- tokenizer/tokenizer_config.json +29 -0
- transformer/config.json +29 -0
- transformer/diffusion_pytorch_model.safetensors +3 -0
- vae/config.json +56 -0
- vae/diffusion_pytorch_model.safetensors +3 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
tokenizer/tokenizer.json filter=lfs diff=lfs merge=lfs -text
|
llm_adapter/config.json
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_class_name": "AnimaLLMAdapter",
|
| 3 |
+
"_diffusers_version": "0.37.0",
|
| 4 |
+
"source_dim": 1024,
|
| 5 |
+
"target_dim": 1024,
|
| 6 |
+
"model_dim": 1024,
|
| 7 |
+
"num_layers": 6,
|
| 8 |
+
"num_heads": 16,
|
| 9 |
+
"mlp_ratio": 4.0,
|
| 10 |
+
"vocab_size": 32128,
|
| 11 |
+
"use_self_attn": true
|
| 12 |
+
}
|
llm_adapter/diffusion_pytorch_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:149d3c0ae9a1b76c5a02a722288a7eadeec306769e2a60f5b34513155c8a2105
|
| 3 |
+
size 269339368
|
llm_adapter/modeling_llm_adapter.py
ADDED
|
@@ -0,0 +1,215 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import torch
|
| 2 |
+
from torch import nn
|
| 3 |
+
import torch.nn.functional as F
|
| 4 |
+
from diffusers.configuration_utils import ConfigMixin, register_to_config
|
| 5 |
+
from diffusers.models.modeling_utils import ModelMixin
|
| 6 |
+
|
| 7 |
+
|
| 8 |
+
def rotate_half(x):
|
| 9 |
+
x1 = x[..., : x.shape[-1] // 2]
|
| 10 |
+
x2 = x[..., x.shape[-1] // 2 :]
|
| 11 |
+
return torch.cat((-x2, x1), dim=-1)
|
| 12 |
+
|
| 13 |
+
|
| 14 |
+
def apply_rotary_pos_emb(x, cos, sin, unsqueeze_dim=1):
|
| 15 |
+
cos = cos.unsqueeze(unsqueeze_dim)
|
| 16 |
+
sin = sin.unsqueeze(unsqueeze_dim)
|
| 17 |
+
return (x * cos) + (rotate_half(x) * sin)
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
class RotaryEmbedding(nn.Module):
|
| 21 |
+
def __init__(self, head_dim):
|
| 22 |
+
super().__init__()
|
| 23 |
+
self.rope_theta = 10000
|
| 24 |
+
inv_freq = 1.0 / (
|
| 25 |
+
self.rope_theta
|
| 26 |
+
** (torch.arange(0, head_dim, 2, dtype=torch.int64).to(dtype=torch.float) / head_dim)
|
| 27 |
+
)
|
| 28 |
+
self.register_buffer("inv_freq", inv_freq, persistent=False)
|
| 29 |
+
|
| 30 |
+
@torch.no_grad()
|
| 31 |
+
def forward(self, x, position_ids):
|
| 32 |
+
inv_freq_expanded = (
|
| 33 |
+
self.inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1).to(x.device)
|
| 34 |
+
)
|
| 35 |
+
position_ids_expanded = position_ids[:, None, :].float()
|
| 36 |
+
|
| 37 |
+
device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
|
| 38 |
+
with torch.autocast(device_type=device_type, enabled=False):
|
| 39 |
+
freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
|
| 40 |
+
emb = torch.cat((freqs, freqs), dim=-1)
|
| 41 |
+
cos = emb.cos()
|
| 42 |
+
sin = emb.sin()
|
| 43 |
+
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
|
| 44 |
+
|
| 45 |
+
|
| 46 |
+
class Attention(nn.Module):
|
| 47 |
+
def __init__(self, query_dim, context_dim, n_heads, head_dim):
|
| 48 |
+
super().__init__()
|
| 49 |
+
inner_dim = head_dim * n_heads
|
| 50 |
+
self.n_heads = n_heads
|
| 51 |
+
self.head_dim = head_dim
|
| 52 |
+
|
| 53 |
+
self.q_proj = nn.Linear(query_dim, inner_dim, bias=False)
|
| 54 |
+
self.q_norm = nn.RMSNorm(head_dim, eps=1e-6)
|
| 55 |
+
self.k_proj = nn.Linear(context_dim, inner_dim, bias=False)
|
| 56 |
+
self.k_norm = nn.RMSNorm(head_dim, eps=1e-6)
|
| 57 |
+
self.v_proj = nn.Linear(context_dim, inner_dim, bias=False)
|
| 58 |
+
self.o_proj = nn.Linear(inner_dim, query_dim, bias=False)
|
| 59 |
+
|
| 60 |
+
def forward(self, x, mask=None, context=None, position_embeddings=None, position_embeddings_context=None):
|
| 61 |
+
context = x if context is None else context
|
| 62 |
+
input_shape = x.shape[:-1]
|
| 63 |
+
q_shape = (*input_shape, self.n_heads, self.head_dim)
|
| 64 |
+
context_shape = context.shape[:-1]
|
| 65 |
+
kv_shape = (*context_shape, self.n_heads, self.head_dim)
|
| 66 |
+
|
| 67 |
+
query_states = self.q_norm(self.q_proj(x).view(q_shape)).transpose(1, 2)
|
| 68 |
+
key_states = self.k_norm(self.k_proj(context).view(kv_shape)).transpose(1, 2)
|
| 69 |
+
value_states = self.v_proj(context).view(kv_shape).transpose(1, 2)
|
| 70 |
+
|
| 71 |
+
if position_embeddings is not None:
|
| 72 |
+
assert position_embeddings_context is not None
|
| 73 |
+
cos, sin = position_embeddings
|
| 74 |
+
query_states = apply_rotary_pos_emb(query_states, cos, sin)
|
| 75 |
+
cos, sin = position_embeddings_context
|
| 76 |
+
key_states = apply_rotary_pos_emb(key_states, cos, sin)
|
| 77 |
+
|
| 78 |
+
attn_output = F.scaled_dot_product_attention(query_states, key_states, value_states, attn_mask=mask)
|
| 79 |
+
attn_output = attn_output.transpose(1, 2).reshape(*input_shape, -1).contiguous()
|
| 80 |
+
return self.o_proj(attn_output)
|
| 81 |
+
|
| 82 |
+
|
| 83 |
+
class TransformerBlock(nn.Module):
|
| 84 |
+
def __init__(self, source_dim, model_dim, num_heads=16, mlp_ratio=4.0, use_self_attn=True):
|
| 85 |
+
super().__init__()
|
| 86 |
+
self.use_self_attn = use_self_attn
|
| 87 |
+
|
| 88 |
+
if self.use_self_attn:
|
| 89 |
+
self.norm_self_attn = nn.RMSNorm(model_dim, eps=1e-6)
|
| 90 |
+
self.self_attn = Attention(
|
| 91 |
+
query_dim=model_dim,
|
| 92 |
+
context_dim=model_dim,
|
| 93 |
+
n_heads=num_heads,
|
| 94 |
+
head_dim=model_dim // num_heads,
|
| 95 |
+
)
|
| 96 |
+
|
| 97 |
+
self.norm_cross_attn = nn.RMSNorm(model_dim, eps=1e-6)
|
| 98 |
+
self.cross_attn = Attention(
|
| 99 |
+
query_dim=model_dim,
|
| 100 |
+
context_dim=source_dim,
|
| 101 |
+
n_heads=num_heads,
|
| 102 |
+
head_dim=model_dim // num_heads,
|
| 103 |
+
)
|
| 104 |
+
|
| 105 |
+
self.norm_mlp = nn.RMSNorm(model_dim, eps=1e-6)
|
| 106 |
+
self.mlp = nn.Sequential(
|
| 107 |
+
nn.Linear(model_dim, int(model_dim * mlp_ratio)),
|
| 108 |
+
nn.GELU(),
|
| 109 |
+
nn.Linear(int(model_dim * mlp_ratio), model_dim),
|
| 110 |
+
)
|
| 111 |
+
|
| 112 |
+
def forward(
|
| 113 |
+
self,
|
| 114 |
+
x,
|
| 115 |
+
context,
|
| 116 |
+
target_attention_mask=None,
|
| 117 |
+
source_attention_mask=None,
|
| 118 |
+
position_embeddings=None,
|
| 119 |
+
position_embeddings_context=None,
|
| 120 |
+
):
|
| 121 |
+
if self.use_self_attn:
|
| 122 |
+
normed = self.norm_self_attn(x)
|
| 123 |
+
attn_out = self.self_attn(
|
| 124 |
+
normed,
|
| 125 |
+
mask=target_attention_mask,
|
| 126 |
+
position_embeddings=position_embeddings,
|
| 127 |
+
position_embeddings_context=position_embeddings,
|
| 128 |
+
)
|
| 129 |
+
x = x + attn_out
|
| 130 |
+
|
| 131 |
+
normed = self.norm_cross_attn(x)
|
| 132 |
+
attn_out = self.cross_attn(
|
| 133 |
+
normed,
|
| 134 |
+
mask=source_attention_mask,
|
| 135 |
+
context=context,
|
| 136 |
+
position_embeddings=position_embeddings,
|
| 137 |
+
position_embeddings_context=position_embeddings_context,
|
| 138 |
+
)
|
| 139 |
+
x = x + attn_out
|
| 140 |
+
x = x + self.mlp(self.norm_mlp(x))
|
| 141 |
+
return x
|
| 142 |
+
|
| 143 |
+
|
| 144 |
+
class AnimaLLMAdapter(ModelMixin, ConfigMixin):
|
| 145 |
+
@register_to_config
|
| 146 |
+
def __init__(
|
| 147 |
+
self,
|
| 148 |
+
source_dim: int = 1024,
|
| 149 |
+
target_dim: int = 1024,
|
| 150 |
+
model_dim: int = 1024,
|
| 151 |
+
num_layers: int = 6,
|
| 152 |
+
num_heads: int = 16,
|
| 153 |
+
mlp_ratio: float = 4.0,
|
| 154 |
+
vocab_size: int = 32128,
|
| 155 |
+
use_self_attn: bool = True,
|
| 156 |
+
):
|
| 157 |
+
super().__init__()
|
| 158 |
+
|
| 159 |
+
self.embed = nn.Embedding(vocab_size, target_dim)
|
| 160 |
+
if model_dim != target_dim:
|
| 161 |
+
self.in_proj = nn.Linear(target_dim, model_dim)
|
| 162 |
+
else:
|
| 163 |
+
self.in_proj = nn.Identity()
|
| 164 |
+
self.rotary_emb = RotaryEmbedding(model_dim // num_heads)
|
| 165 |
+
self.blocks = nn.ModuleList(
|
| 166 |
+
[
|
| 167 |
+
TransformerBlock(
|
| 168 |
+
source_dim,
|
| 169 |
+
model_dim,
|
| 170 |
+
num_heads=num_heads,
|
| 171 |
+
mlp_ratio=mlp_ratio,
|
| 172 |
+
use_self_attn=use_self_attn,
|
| 173 |
+
)
|
| 174 |
+
for _ in range(num_layers)
|
| 175 |
+
]
|
| 176 |
+
)
|
| 177 |
+
self.out_proj = nn.Linear(model_dim, target_dim)
|
| 178 |
+
self.norm = nn.RMSNorm(target_dim, eps=1e-6)
|
| 179 |
+
|
| 180 |
+
def forward(
|
| 181 |
+
self,
|
| 182 |
+
source_hidden_states: torch.Tensor,
|
| 183 |
+
target_input_ids: torch.Tensor,
|
| 184 |
+
target_attention_mask: torch.Tensor = None,
|
| 185 |
+
source_attention_mask: torch.Tensor = None,
|
| 186 |
+
) -> torch.Tensor:
|
| 187 |
+
if target_attention_mask is not None:
|
| 188 |
+
target_attention_mask = target_attention_mask.to(torch.bool)
|
| 189 |
+
if target_attention_mask.ndim == 2:
|
| 190 |
+
target_attention_mask = target_attention_mask.unsqueeze(1).unsqueeze(1)
|
| 191 |
+
|
| 192 |
+
if source_attention_mask is not None:
|
| 193 |
+
source_attention_mask = source_attention_mask.to(torch.bool)
|
| 194 |
+
if source_attention_mask.ndim == 2:
|
| 195 |
+
source_attention_mask = source_attention_mask.unsqueeze(1).unsqueeze(1)
|
| 196 |
+
|
| 197 |
+
x = self.in_proj(self.embed(target_input_ids))
|
| 198 |
+
context = source_hidden_states
|
| 199 |
+
|
| 200 |
+
position_ids = torch.arange(x.shape[1], device=x.device).unsqueeze(0)
|
| 201 |
+
position_ids_context = torch.arange(context.shape[1], device=x.device).unsqueeze(0)
|
| 202 |
+
position_embeddings = self.rotary_emb(x, position_ids)
|
| 203 |
+
position_embeddings_context = self.rotary_emb(x, position_ids_context)
|
| 204 |
+
|
| 205 |
+
for block in self.blocks:
|
| 206 |
+
x = block(
|
| 207 |
+
x,
|
| 208 |
+
context,
|
| 209 |
+
target_attention_mask=target_attention_mask,
|
| 210 |
+
source_attention_mask=source_attention_mask,
|
| 211 |
+
position_embeddings=position_embeddings,
|
| 212 |
+
position_embeddings_context=position_embeddings_context,
|
| 213 |
+
)
|
| 214 |
+
|
| 215 |
+
return self.norm(self.out_proj(x))
|
model_index.json
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_class_name": [
|
| 3 |
+
"pipeline",
|
| 4 |
+
"AnimaTextToImagePipeline"
|
| 5 |
+
],
|
| 6 |
+
"_diffusers_version": "0.37.0",
|
| 7 |
+
"text_encoder": [
|
| 8 |
+
"transformers",
|
| 9 |
+
"Qwen3Model"
|
| 10 |
+
],
|
| 11 |
+
"tokenizer": [
|
| 12 |
+
"transformers",
|
| 13 |
+
"PreTrainedTokenizerFast"
|
| 14 |
+
],
|
| 15 |
+
"t5_tokenizer": [
|
| 16 |
+
"transformers",
|
| 17 |
+
"T5TokenizerFast"
|
| 18 |
+
],
|
| 19 |
+
"llm_adapter": [
|
| 20 |
+
"modeling_llm_adapter",
|
| 21 |
+
"AnimaLLMAdapter"
|
| 22 |
+
],
|
| 23 |
+
"transformer": [
|
| 24 |
+
"diffusers",
|
| 25 |
+
"CosmosTransformer3DModel"
|
| 26 |
+
],
|
| 27 |
+
"vae": [
|
| 28 |
+
"diffusers",
|
| 29 |
+
"AutoencoderKLWan"
|
| 30 |
+
],
|
| 31 |
+
"scheduler": [
|
| 32 |
+
"diffusers",
|
| 33 |
+
"FlowMatchEulerDiscreteScheduler"
|
| 34 |
+
]
|
| 35 |
+
}
|
pipeline.py
ADDED
|
@@ -0,0 +1,371 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import inspect
|
| 2 |
+
from typing import Callable, Dict, List, Optional, Union
|
| 3 |
+
|
| 4 |
+
import numpy as np
|
| 5 |
+
import torch
|
| 6 |
+
from transformers import PreTrainedModel, PreTrainedTokenizerFast
|
| 7 |
+
|
| 8 |
+
from diffusers.callbacks import MultiPipelineCallbacks, PipelineCallback
|
| 9 |
+
from diffusers.models import AutoencoderKLWan, CosmosTransformer3DModel
|
| 10 |
+
from diffusers.schedulers import FlowMatchEulerDiscreteScheduler
|
| 11 |
+
from diffusers.utils import logging
|
| 12 |
+
from diffusers.utils.torch_utils import randn_tensor
|
| 13 |
+
from diffusers.video_processor import VideoProcessor
|
| 14 |
+
from diffusers.pipelines.pipeline_utils import DiffusionPipeline
|
| 15 |
+
from diffusers.pipelines.cosmos.pipeline_output import CosmosImagePipelineOutput
|
| 16 |
+
|
| 17 |
+
logger = logging.get_logger(__name__)
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
def retrieve_timesteps(scheduler, num_inference_steps=None, device=None, timesteps=None, sigmas=None, **kwargs):
|
| 21 |
+
if timesteps is not None and sigmas is not None:
|
| 22 |
+
raise ValueError("Only one of `timesteps` or `sigmas` can be passed.")
|
| 23 |
+
if timesteps is not None:
|
| 24 |
+
scheduler.set_timesteps(timesteps=timesteps, device=device, **kwargs)
|
| 25 |
+
timesteps = scheduler.timesteps
|
| 26 |
+
num_inference_steps = len(timesteps)
|
| 27 |
+
elif sigmas is not None:
|
| 28 |
+
scheduler.set_timesteps(sigmas=sigmas, device=device, **kwargs)
|
| 29 |
+
timesteps = scheduler.timesteps
|
| 30 |
+
num_inference_steps = len(timesteps)
|
| 31 |
+
else:
|
| 32 |
+
scheduler.set_timesteps(num_inference_steps, device=device, **kwargs)
|
| 33 |
+
timesteps = scheduler.timesteps
|
| 34 |
+
return timesteps, num_inference_steps
|
| 35 |
+
|
| 36 |
+
|
| 37 |
+
class AnimaTextToImagePipeline(DiffusionPipeline):
|
| 38 |
+
"""Pipeline for text-to-image generation using the Anima model.
|
| 39 |
+
|
| 40 |
+
Anima uses a Cosmos Predict2 backbone with a Qwen3 text encoder and an LLM adapter
|
| 41 |
+
that cross-attends T5 token embeddings to Qwen3 hidden states.
|
| 42 |
+
"""
|
| 43 |
+
|
| 44 |
+
model_cpu_offload_seq = "text_encoder->llm_adapter->transformer->vae"
|
| 45 |
+
_callback_tensor_inputs = ["latents", "prompt_embeds", "negative_prompt_embeds"]
|
| 46 |
+
|
| 47 |
+
def __init__(
|
| 48 |
+
self,
|
| 49 |
+
text_encoder: PreTrainedModel,
|
| 50 |
+
tokenizer: PreTrainedTokenizerFast,
|
| 51 |
+
t5_tokenizer: PreTrainedTokenizerFast,
|
| 52 |
+
llm_adapter,
|
| 53 |
+
transformer: CosmosTransformer3DModel,
|
| 54 |
+
vae: AutoencoderKLWan,
|
| 55 |
+
scheduler: FlowMatchEulerDiscreteScheduler,
|
| 56 |
+
):
|
| 57 |
+
super().__init__()
|
| 58 |
+
|
| 59 |
+
self.register_modules(
|
| 60 |
+
text_encoder=text_encoder,
|
| 61 |
+
tokenizer=tokenizer,
|
| 62 |
+
t5_tokenizer=t5_tokenizer,
|
| 63 |
+
llm_adapter=llm_adapter,
|
| 64 |
+
transformer=transformer,
|
| 65 |
+
vae=vae,
|
| 66 |
+
scheduler=scheduler,
|
| 67 |
+
)
|
| 68 |
+
|
| 69 |
+
self.vae_scale_factor_temporal = 2 ** sum(self.vae.temperal_downsample) if getattr(self, "vae", None) else 4
|
| 70 |
+
self.vae_scale_factor_spatial = 2 ** len(self.vae.temperal_downsample) if getattr(self, "vae", None) else 8
|
| 71 |
+
self.video_processor = VideoProcessor(vae_scale_factor=self.vae_scale_factor_spatial)
|
| 72 |
+
|
| 73 |
+
def _encode_prompt(
|
| 74 |
+
self,
|
| 75 |
+
prompt: Union[str, List[str]],
|
| 76 |
+
device: torch.device,
|
| 77 |
+
dtype: torch.dtype,
|
| 78 |
+
max_sequence_length: int = 512,
|
| 79 |
+
):
|
| 80 |
+
"""Encode prompt through Qwen3 and run LLM adapter with T5 token IDs."""
|
| 81 |
+
prompt = [prompt] if isinstance(prompt, str) else prompt
|
| 82 |
+
batch_size = len(prompt)
|
| 83 |
+
|
| 84 |
+
# Check for empty prompts - return zero embeddings directly
|
| 85 |
+
all_empty = all(p.strip() == "" for p in prompt)
|
| 86 |
+
if all_empty:
|
| 87 |
+
return torch.zeros(batch_size, 512, self.llm_adapter.config.target_dim, device=device, dtype=dtype)
|
| 88 |
+
|
| 89 |
+
# Tokenize with Qwen3 tokenizer
|
| 90 |
+
qwen_inputs = self.tokenizer(
|
| 91 |
+
prompt,
|
| 92 |
+
padding=True,
|
| 93 |
+
truncation=True,
|
| 94 |
+
max_length=max_sequence_length,
|
| 95 |
+
return_tensors="pt",
|
| 96 |
+
)
|
| 97 |
+
qwen_input_ids = qwen_inputs.input_ids.to(device)
|
| 98 |
+
qwen_attention_mask = qwen_inputs.attention_mask.to(device)
|
| 99 |
+
|
| 100 |
+
# Get Qwen3 hidden states
|
| 101 |
+
qwen_outputs = self.text_encoder(
|
| 102 |
+
input_ids=qwen_input_ids,
|
| 103 |
+
attention_mask=qwen_attention_mask,
|
| 104 |
+
)
|
| 105 |
+
qwen_hidden_states = qwen_outputs.last_hidden_state.to(dtype=dtype)
|
| 106 |
+
|
| 107 |
+
# Tokenize with T5 tokenizer (we only need the IDs for the adapter embedding)
|
| 108 |
+
t5_inputs = self.t5_tokenizer(
|
| 109 |
+
prompt,
|
| 110 |
+
padding=True,
|
| 111 |
+
truncation=True,
|
| 112 |
+
max_length=max_sequence_length,
|
| 113 |
+
return_tensors="pt",
|
| 114 |
+
)
|
| 115 |
+
t5_input_ids = t5_inputs.input_ids.to(device)
|
| 116 |
+
|
| 117 |
+
# Run LLM adapter: T5 token embeddings attend to Qwen3 hidden states
|
| 118 |
+
adapted_embeds = self.llm_adapter(
|
| 119 |
+
source_hidden_states=qwen_hidden_states,
|
| 120 |
+
target_input_ids=t5_input_ids,
|
| 121 |
+
)
|
| 122 |
+
|
| 123 |
+
# Pad to 512 sequence length if shorter
|
| 124 |
+
if adapted_embeds.shape[1] < 512:
|
| 125 |
+
adapted_embeds = torch.nn.functional.pad(
|
| 126 |
+
adapted_embeds, (0, 0, 0, 512 - adapted_embeds.shape[1])
|
| 127 |
+
)
|
| 128 |
+
|
| 129 |
+
return adapted_embeds
|
| 130 |
+
|
| 131 |
+
def encode_prompt(
|
| 132 |
+
self,
|
| 133 |
+
prompt: Union[str, List[str]],
|
| 134 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
| 135 |
+
do_classifier_free_guidance: bool = True,
|
| 136 |
+
num_images_per_prompt: int = 1,
|
| 137 |
+
prompt_embeds: Optional[torch.Tensor] = None,
|
| 138 |
+
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
| 139 |
+
max_sequence_length: int = 512,
|
| 140 |
+
device: Optional[torch.device] = None,
|
| 141 |
+
dtype: Optional[torch.dtype] = None,
|
| 142 |
+
):
|
| 143 |
+
device = device or self._execution_device
|
| 144 |
+
dtype = dtype or self.text_encoder.dtype
|
| 145 |
+
prompt = [prompt] if isinstance(prompt, str) else prompt
|
| 146 |
+
|
| 147 |
+
if prompt is not None:
|
| 148 |
+
batch_size = len(prompt)
|
| 149 |
+
else:
|
| 150 |
+
batch_size = prompt_embeds.shape[0]
|
| 151 |
+
|
| 152 |
+
if prompt_embeds is None:
|
| 153 |
+
prompt_embeds = self._encode_prompt(prompt, device, dtype, max_sequence_length)
|
| 154 |
+
_, seq_len, _ = prompt_embeds.shape
|
| 155 |
+
prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
| 156 |
+
prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
|
| 157 |
+
|
| 158 |
+
if do_classifier_free_guidance and negative_prompt_embeds is None:
|
| 159 |
+
negative_prompt = negative_prompt or ""
|
| 160 |
+
negative_prompt = batch_size * [negative_prompt] if isinstance(negative_prompt, str) else negative_prompt
|
| 161 |
+
negative_prompt_embeds = self._encode_prompt(negative_prompt, device, dtype, max_sequence_length)
|
| 162 |
+
_, seq_len, _ = negative_prompt_embeds.shape
|
| 163 |
+
negative_prompt_embeds = negative_prompt_embeds.repeat(1, num_images_per_prompt, 1)
|
| 164 |
+
negative_prompt_embeds = negative_prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)
|
| 165 |
+
|
| 166 |
+
return prompt_embeds, negative_prompt_embeds
|
| 167 |
+
|
| 168 |
+
def prepare_latents(
|
| 169 |
+
self,
|
| 170 |
+
batch_size: int,
|
| 171 |
+
num_channels_latents: int,
|
| 172 |
+
height: int,
|
| 173 |
+
width: int,
|
| 174 |
+
num_frames: int = 1,
|
| 175 |
+
dtype: torch.dtype = None,
|
| 176 |
+
device: torch.device = None,
|
| 177 |
+
generator=None,
|
| 178 |
+
latents: torch.Tensor = None,
|
| 179 |
+
):
|
| 180 |
+
num_latent_frames = (num_frames - 1) // self.vae_scale_factor_temporal + 1
|
| 181 |
+
latent_height = height // self.vae_scale_factor_spatial
|
| 182 |
+
latent_width = width // self.vae_scale_factor_spatial
|
| 183 |
+
|
| 184 |
+
if latents is not None:
|
| 185 |
+
return latents.to(device=device, dtype=dtype)
|
| 186 |
+
|
| 187 |
+
shape = (batch_size, num_channels_latents, num_latent_frames, latent_height, latent_width)
|
| 188 |
+
latents = randn_tensor(shape, generator=generator, device=device, dtype=dtype)
|
| 189 |
+
return latents
|
| 190 |
+
|
| 191 |
+
def check_inputs(self, prompt, height, width, prompt_embeds=None):
|
| 192 |
+
if height % 16 != 0 or width % 16 != 0:
|
| 193 |
+
raise ValueError(f"`height` and `width` have to be divisible by 16 but are {height} and {width}.")
|
| 194 |
+
if prompt is not None and prompt_embeds is not None:
|
| 195 |
+
raise ValueError("Cannot forward both `prompt` and `prompt_embeds`.")
|
| 196 |
+
elif prompt is None and prompt_embeds is None:
|
| 197 |
+
raise ValueError("Provide either `prompt` or `prompt_embeds`.")
|
| 198 |
+
|
| 199 |
+
@property
|
| 200 |
+
def guidance_scale(self):
|
| 201 |
+
return self._guidance_scale
|
| 202 |
+
|
| 203 |
+
@property
|
| 204 |
+
def do_classifier_free_guidance(self):
|
| 205 |
+
return self._guidance_scale > 1.0
|
| 206 |
+
|
| 207 |
+
@property
|
| 208 |
+
def num_timesteps(self):
|
| 209 |
+
return self._num_timesteps
|
| 210 |
+
|
| 211 |
+
@property
|
| 212 |
+
def interrupt(self):
|
| 213 |
+
return self._interrupt
|
| 214 |
+
|
| 215 |
+
@torch.no_grad()
|
| 216 |
+
def __call__(
|
| 217 |
+
self,
|
| 218 |
+
prompt: Union[str, List[str]] = None,
|
| 219 |
+
negative_prompt: Optional[Union[str, List[str]]] = None,
|
| 220 |
+
height: int = 768,
|
| 221 |
+
width: int = 1360,
|
| 222 |
+
num_inference_steps: int = 35,
|
| 223 |
+
guidance_scale: float = 7.0,
|
| 224 |
+
num_images_per_prompt: Optional[int] = 1,
|
| 225 |
+
generator: Optional[Union[torch.Generator, List[torch.Generator]]] = None,
|
| 226 |
+
latents: Optional[torch.Tensor] = None,
|
| 227 |
+
prompt_embeds: Optional[torch.Tensor] = None,
|
| 228 |
+
negative_prompt_embeds: Optional[torch.Tensor] = None,
|
| 229 |
+
output_type: Optional[str] = "pil",
|
| 230 |
+
return_dict: bool = True,
|
| 231 |
+
callback_on_step_end: Optional[
|
| 232 |
+
Union[Callable[[int, int, Dict], None], PipelineCallback, MultiPipelineCallbacks]
|
| 233 |
+
] = None,
|
| 234 |
+
callback_on_step_end_tensor_inputs: List[str] = ["latents"],
|
| 235 |
+
max_sequence_length: int = 512,
|
| 236 |
+
):
|
| 237 |
+
if isinstance(callback_on_step_end, (PipelineCallback, MultiPipelineCallbacks)):
|
| 238 |
+
callback_on_step_end_tensor_inputs = callback_on_step_end.tensor_inputs
|
| 239 |
+
|
| 240 |
+
num_frames = 1
|
| 241 |
+
|
| 242 |
+
self.check_inputs(prompt, height, width, prompt_embeds)
|
| 243 |
+
self._guidance_scale = guidance_scale
|
| 244 |
+
self._current_timestep = None
|
| 245 |
+
self._interrupt = False
|
| 246 |
+
|
| 247 |
+
device = self._execution_device
|
| 248 |
+
|
| 249 |
+
if prompt is not None and isinstance(prompt, str):
|
| 250 |
+
batch_size = 1
|
| 251 |
+
elif prompt is not None and isinstance(prompt, list):
|
| 252 |
+
batch_size = len(prompt)
|
| 253 |
+
else:
|
| 254 |
+
batch_size = prompt_embeds.shape[0]
|
| 255 |
+
|
| 256 |
+
# Encode prompt
|
| 257 |
+
prompt_embeds, negative_prompt_embeds = self.encode_prompt(
|
| 258 |
+
prompt=prompt,
|
| 259 |
+
negative_prompt=negative_prompt,
|
| 260 |
+
do_classifier_free_guidance=self.do_classifier_free_guidance,
|
| 261 |
+
num_images_per_prompt=num_images_per_prompt,
|
| 262 |
+
prompt_embeds=prompt_embeds,
|
| 263 |
+
negative_prompt_embeds=negative_prompt_embeds,
|
| 264 |
+
device=device,
|
| 265 |
+
max_sequence_length=max_sequence_length,
|
| 266 |
+
)
|
| 267 |
+
|
| 268 |
+
# Prepare timesteps - use default descending schedule (1→0)
|
| 269 |
+
timesteps, num_inference_steps = retrieve_timesteps(
|
| 270 |
+
self.scheduler, num_inference_steps=num_inference_steps, device=device
|
| 271 |
+
)
|
| 272 |
+
|
| 273 |
+
# Prepare latents
|
| 274 |
+
transformer_dtype = self.transformer.dtype
|
| 275 |
+
num_channels_latents = self.transformer.config.in_channels
|
| 276 |
+
latents = self.prepare_latents(
|
| 277 |
+
batch_size * num_images_per_prompt,
|
| 278 |
+
num_channels_latents,
|
| 279 |
+
height,
|
| 280 |
+
width,
|
| 281 |
+
num_frames,
|
| 282 |
+
torch.float32,
|
| 283 |
+
device,
|
| 284 |
+
generator,
|
| 285 |
+
latents,
|
| 286 |
+
)
|
| 287 |
+
|
| 288 |
+
padding_mask = latents.new_zeros(1, 1, height, width, dtype=transformer_dtype)
|
| 289 |
+
|
| 290 |
+
# Denoising loop using CONST preconditioning (flow matching velocity model):
|
| 291 |
+
# - c_in = 1.0 (no input scaling)
|
| 292 |
+
# - timestep = sigma (passed directly)
|
| 293 |
+
# - model output is the velocity: denoised = x - velocity * sigma
|
| 294 |
+
# - CFG applied to velocity (equivalent to applying to denoised for linear preconditioning)
|
| 295 |
+
num_warmup_steps = len(timesteps) - num_inference_steps * self.scheduler.order
|
| 296 |
+
self._num_timesteps = len(timesteps)
|
| 297 |
+
|
| 298 |
+
with self.progress_bar(total=num_inference_steps) as progress_bar:
|
| 299 |
+
for i, t in enumerate(timesteps):
|
| 300 |
+
if self.interrupt:
|
| 301 |
+
continue
|
| 302 |
+
|
| 303 |
+
self._current_timestep = t
|
| 304 |
+
sigma = self.scheduler.sigmas[i]
|
| 305 |
+
|
| 306 |
+
# Pass sigma directly as timestep (CONST preconditioning)
|
| 307 |
+
timestep = sigma.expand(latents.shape[0]).to(transformer_dtype)
|
| 308 |
+
latent_model_input = latents.to(transformer_dtype)
|
| 309 |
+
|
| 310 |
+
# Model predicts velocity (raw output IS the velocity for CONST)
|
| 311 |
+
velocity = self.transformer(
|
| 312 |
+
hidden_states=latent_model_input,
|
| 313 |
+
timestep=timestep,
|
| 314 |
+
encoder_hidden_states=prompt_embeds,
|
| 315 |
+
padding_mask=padding_mask,
|
| 316 |
+
return_dict=False,
|
| 317 |
+
)[0].float()
|
| 318 |
+
|
| 319 |
+
if self.do_classifier_free_guidance:
|
| 320 |
+
velocity_uncond = self.transformer(
|
| 321 |
+
hidden_states=latent_model_input,
|
| 322 |
+
timestep=timestep,
|
| 323 |
+
encoder_hidden_states=negative_prompt_embeds,
|
| 324 |
+
padding_mask=padding_mask,
|
| 325 |
+
return_dict=False,
|
| 326 |
+
)[0].float()
|
| 327 |
+
velocity = velocity_uncond + self.guidance_scale * (velocity - velocity_uncond)
|
| 328 |
+
|
| 329 |
+
# Euler step: scheduler computes x_next = x + (sigma_next - sigma) * velocity
|
| 330 |
+
latents = self.scheduler.step(velocity, t, latents, return_dict=False)[0]
|
| 331 |
+
|
| 332 |
+
if callback_on_step_end is not None:
|
| 333 |
+
callback_kwargs = {}
|
| 334 |
+
for k in callback_on_step_end_tensor_inputs:
|
| 335 |
+
callback_kwargs[k] = locals()[k]
|
| 336 |
+
callback_outputs = callback_on_step_end(self, i, t, callback_kwargs)
|
| 337 |
+
latents = callback_outputs.pop("latents", latents)
|
| 338 |
+
prompt_embeds = callback_outputs.pop("prompt_embeds", prompt_embeds)
|
| 339 |
+
negative_prompt_embeds = callback_outputs.pop("negative_prompt_embeds", negative_prompt_embeds)
|
| 340 |
+
|
| 341 |
+
if i == len(timesteps) - 1 or ((i + 1) > num_warmup_steps and (i + 1) % self.scheduler.order == 0):
|
| 342 |
+
progress_bar.update()
|
| 343 |
+
|
| 344 |
+
self._current_timestep = None
|
| 345 |
+
|
| 346 |
+
if not output_type == "latent":
|
| 347 |
+
latents_mean = (
|
| 348 |
+
torch.tensor(self.vae.config.latents_mean)
|
| 349 |
+
.view(1, self.vae.config.z_dim, 1, 1, 1)
|
| 350 |
+
.to(latents.device, latents.dtype)
|
| 351 |
+
)
|
| 352 |
+
latents_std = 1.0 / torch.tensor(self.vae.config.latents_std).view(1, self.vae.config.z_dim, 1, 1, 1).to(
|
| 353 |
+
latents.device, latents.dtype
|
| 354 |
+
)
|
| 355 |
+
latents = latents / latents_std + latents_mean
|
| 356 |
+
video = self.vae.decode(latents.to(self.vae.dtype), return_dict=False)[0]
|
| 357 |
+
video = self.video_processor.postprocess_video(video, output_type=output_type)
|
| 358 |
+
image = [batch[0] for batch in video]
|
| 359 |
+
if isinstance(video, torch.Tensor):
|
| 360 |
+
image = torch.stack(image)
|
| 361 |
+
elif isinstance(video, np.ndarray):
|
| 362 |
+
image = np.stack(image)
|
| 363 |
+
else:
|
| 364 |
+
image = latents[:, :, 0]
|
| 365 |
+
|
| 366 |
+
self.maybe_free_model_hooks()
|
| 367 |
+
|
| 368 |
+
if not return_dict:
|
| 369 |
+
return (image,)
|
| 370 |
+
|
| 371 |
+
return CosmosImagePipelineOutput(images=image)
|
scheduler/scheduler_config.json
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_class_name": "FlowMatchEulerDiscreteScheduler",
|
| 3 |
+
"_diffusers_version": "0.37.0",
|
| 4 |
+
"num_train_timesteps": 1000,
|
| 5 |
+
"shift": 3.0
|
| 6 |
+
}
|
t5_tokenizer/tokenizer.json
ADDED
|
The diff for this file is too large to render.
See raw diff
|
|
|
t5_tokenizer/tokenizer_config.json
ADDED
|
@@ -0,0 +1,113 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"backend": "tokenizers",
|
| 3 |
+
"clean_up_tokenization_spaces": true,
|
| 4 |
+
"eos_token": "</s>",
|
| 5 |
+
"extra_ids": 100,
|
| 6 |
+
"extra_special_tokens": [
|
| 7 |
+
"<extra_id_0>",
|
| 8 |
+
"<extra_id_1>",
|
| 9 |
+
"<extra_id_2>",
|
| 10 |
+
"<extra_id_3>",
|
| 11 |
+
"<extra_id_4>",
|
| 12 |
+
"<extra_id_5>",
|
| 13 |
+
"<extra_id_6>",
|
| 14 |
+
"<extra_id_7>",
|
| 15 |
+
"<extra_id_8>",
|
| 16 |
+
"<extra_id_9>",
|
| 17 |
+
"<extra_id_10>",
|
| 18 |
+
"<extra_id_11>",
|
| 19 |
+
"<extra_id_12>",
|
| 20 |
+
"<extra_id_13>",
|
| 21 |
+
"<extra_id_14>",
|
| 22 |
+
"<extra_id_15>",
|
| 23 |
+
"<extra_id_16>",
|
| 24 |
+
"<extra_id_17>",
|
| 25 |
+
"<extra_id_18>",
|
| 26 |
+
"<extra_id_19>",
|
| 27 |
+
"<extra_id_20>",
|
| 28 |
+
"<extra_id_21>",
|
| 29 |
+
"<extra_id_22>",
|
| 30 |
+
"<extra_id_23>",
|
| 31 |
+
"<extra_id_24>",
|
| 32 |
+
"<extra_id_25>",
|
| 33 |
+
"<extra_id_26>",
|
| 34 |
+
"<extra_id_27>",
|
| 35 |
+
"<extra_id_28>",
|
| 36 |
+
"<extra_id_29>",
|
| 37 |
+
"<extra_id_30>",
|
| 38 |
+
"<extra_id_31>",
|
| 39 |
+
"<extra_id_32>",
|
| 40 |
+
"<extra_id_33>",
|
| 41 |
+
"<extra_id_34>",
|
| 42 |
+
"<extra_id_35>",
|
| 43 |
+
"<extra_id_36>",
|
| 44 |
+
"<extra_id_37>",
|
| 45 |
+
"<extra_id_38>",
|
| 46 |
+
"<extra_id_39>",
|
| 47 |
+
"<extra_id_40>",
|
| 48 |
+
"<extra_id_41>",
|
| 49 |
+
"<extra_id_42>",
|
| 50 |
+
"<extra_id_43>",
|
| 51 |
+
"<extra_id_44>",
|
| 52 |
+
"<extra_id_45>",
|
| 53 |
+
"<extra_id_46>",
|
| 54 |
+
"<extra_id_47>",
|
| 55 |
+
"<extra_id_48>",
|
| 56 |
+
"<extra_id_49>",
|
| 57 |
+
"<extra_id_50>",
|
| 58 |
+
"<extra_id_51>",
|
| 59 |
+
"<extra_id_52>",
|
| 60 |
+
"<extra_id_53>",
|
| 61 |
+
"<extra_id_54>",
|
| 62 |
+
"<extra_id_55>",
|
| 63 |
+
"<extra_id_56>",
|
| 64 |
+
"<extra_id_57>",
|
| 65 |
+
"<extra_id_58>",
|
| 66 |
+
"<extra_id_59>",
|
| 67 |
+
"<extra_id_60>",
|
| 68 |
+
"<extra_id_61>",
|
| 69 |
+
"<extra_id_62>",
|
| 70 |
+
"<extra_id_63>",
|
| 71 |
+
"<extra_id_64>",
|
| 72 |
+
"<extra_id_65>",
|
| 73 |
+
"<extra_id_66>",
|
| 74 |
+
"<extra_id_67>",
|
| 75 |
+
"<extra_id_68>",
|
| 76 |
+
"<extra_id_69>",
|
| 77 |
+
"<extra_id_70>",
|
| 78 |
+
"<extra_id_71>",
|
| 79 |
+
"<extra_id_72>",
|
| 80 |
+
"<extra_id_73>",
|
| 81 |
+
"<extra_id_74>",
|
| 82 |
+
"<extra_id_75>",
|
| 83 |
+
"<extra_id_76>",
|
| 84 |
+
"<extra_id_77>",
|
| 85 |
+
"<extra_id_78>",
|
| 86 |
+
"<extra_id_79>",
|
| 87 |
+
"<extra_id_80>",
|
| 88 |
+
"<extra_id_81>",
|
| 89 |
+
"<extra_id_82>",
|
| 90 |
+
"<extra_id_83>",
|
| 91 |
+
"<extra_id_84>",
|
| 92 |
+
"<extra_id_85>",
|
| 93 |
+
"<extra_id_86>",
|
| 94 |
+
"<extra_id_87>",
|
| 95 |
+
"<extra_id_88>",
|
| 96 |
+
"<extra_id_89>",
|
| 97 |
+
"<extra_id_90>",
|
| 98 |
+
"<extra_id_91>",
|
| 99 |
+
"<extra_id_92>",
|
| 100 |
+
"<extra_id_93>",
|
| 101 |
+
"<extra_id_94>",
|
| 102 |
+
"<extra_id_95>",
|
| 103 |
+
"<extra_id_96>",
|
| 104 |
+
"<extra_id_97>",
|
| 105 |
+
"<extra_id_98>",
|
| 106 |
+
"<extra_id_99>"
|
| 107 |
+
],
|
| 108 |
+
"is_local": false,
|
| 109 |
+
"model_max_length": 512,
|
| 110 |
+
"pad_token": "<pad>",
|
| 111 |
+
"tokenizer_class": "T5Tokenizer",
|
| 112 |
+
"unk_token": "<unk>"
|
| 113 |
+
}
|
text_encoder/config.json
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"Qwen3Model"
|
| 4 |
+
],
|
| 5 |
+
"model_type": "qwen3",
|
| 6 |
+
"vocab_size": 151936,
|
| 7 |
+
"hidden_size": 1024,
|
| 8 |
+
"intermediate_size": 3072,
|
| 9 |
+
"num_hidden_layers": 28,
|
| 10 |
+
"num_attention_heads": 16,
|
| 11 |
+
"num_key_value_heads": 8,
|
| 12 |
+
"head_dim": 128,
|
| 13 |
+
"hidden_act": "silu",
|
| 14 |
+
"max_position_embeddings": 32768,
|
| 15 |
+
"rms_norm_eps": 1e-06,
|
| 16 |
+
"rope_theta": 1000000.0,
|
| 17 |
+
"attention_bias": false,
|
| 18 |
+
"attention_dropout": 0.0,
|
| 19 |
+
"use_cache": false,
|
| 20 |
+
"tie_word_embeddings": false,
|
| 21 |
+
"torch_dtype": "bfloat16"
|
| 22 |
+
}
|
text_encoder/model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:d10aa56a4da8a95d954d99228d9e20e27f96ac5fc8aa41b89a41532b16bb4817
|
| 3 |
+
size 1192135064
|
tokenizer/chat_template.jinja
ADDED
|
@@ -0,0 +1,89 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{%- if tools %}
|
| 2 |
+
{{- '<|im_start|>system\n' }}
|
| 3 |
+
{%- if messages[0].role == 'system' %}
|
| 4 |
+
{{- messages[0].content + '\n\n' }}
|
| 5 |
+
{%- endif %}
|
| 6 |
+
{{- "# Tools\n\nYou may call one or more functions to assist with the user query.\n\nYou are provided with function signatures within <tools></tools> XML tags:\n<tools>" }}
|
| 7 |
+
{%- for tool in tools %}
|
| 8 |
+
{{- "\n" }}
|
| 9 |
+
{{- tool | tojson }}
|
| 10 |
+
{%- endfor %}
|
| 11 |
+
{{- "\n</tools>\n\nFor each function call, return a json object with function name and arguments within <tool_call></tool_call> XML tags:\n<tool_call>\n{\"name\": <function-name>, \"arguments\": <args-json-object>}\n</tool_call><|im_end|>\n" }}
|
| 12 |
+
{%- else %}
|
| 13 |
+
{%- if messages[0].role == 'system' %}
|
| 14 |
+
{{- '<|im_start|>system\n' + messages[0].content + '<|im_end|>\n' }}
|
| 15 |
+
{%- endif %}
|
| 16 |
+
{%- endif %}
|
| 17 |
+
{%- set ns = namespace(multi_step_tool=true, last_query_index=messages|length - 1) %}
|
| 18 |
+
{%- for message in messages[::-1] %}
|
| 19 |
+
{%- set index = (messages|length - 1) - loop.index0 %}
|
| 20 |
+
{%- if ns.multi_step_tool and message.role == "user" and message.content is string and not(message.content.startswith('<tool_response>') and message.content.endswith('</tool_response>')) %}
|
| 21 |
+
{%- set ns.multi_step_tool = false %}
|
| 22 |
+
{%- set ns.last_query_index = index %}
|
| 23 |
+
{%- endif %}
|
| 24 |
+
{%- endfor %}
|
| 25 |
+
{%- for message in messages %}
|
| 26 |
+
{%- if message.content is string %}
|
| 27 |
+
{%- set content = message.content %}
|
| 28 |
+
{%- else %}
|
| 29 |
+
{%- set content = '' %}
|
| 30 |
+
{%- endif %}
|
| 31 |
+
{%- if (message.role == "user") or (message.role == "system" and not loop.first) %}
|
| 32 |
+
{{- '<|im_start|>' + message.role + '\n' + content + '<|im_end|>' + '\n' }}
|
| 33 |
+
{%- elif message.role == "assistant" %}
|
| 34 |
+
{%- set reasoning_content = '' %}
|
| 35 |
+
{%- if message.reasoning_content is string %}
|
| 36 |
+
{%- set reasoning_content = message.reasoning_content %}
|
| 37 |
+
{%- else %}
|
| 38 |
+
{%- if '</think>' in content %}
|
| 39 |
+
{%- set reasoning_content = content.split('</think>')[0].rstrip('\n').split('<think>')[-1].lstrip('\n') %}
|
| 40 |
+
{%- set content = content.split('</think>')[-1].lstrip('\n') %}
|
| 41 |
+
{%- endif %}
|
| 42 |
+
{%- endif %}
|
| 43 |
+
{%- if loop.index0 > ns.last_query_index %}
|
| 44 |
+
{%- if loop.last or (not loop.last and reasoning_content) %}
|
| 45 |
+
{{- '<|im_start|>' + message.role + '\n<think>\n' + reasoning_content.strip('\n') + '\n</think>\n\n' + content.lstrip('\n') }}
|
| 46 |
+
{%- else %}
|
| 47 |
+
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 48 |
+
{%- endif %}
|
| 49 |
+
{%- else %}
|
| 50 |
+
{{- '<|im_start|>' + message.role + '\n' + content }}
|
| 51 |
+
{%- endif %}
|
| 52 |
+
{%- if message.tool_calls %}
|
| 53 |
+
{%- for tool_call in message.tool_calls %}
|
| 54 |
+
{%- if (loop.first and content) or (not loop.first) %}
|
| 55 |
+
{{- '\n' }}
|
| 56 |
+
{%- endif %}
|
| 57 |
+
{%- if tool_call.function %}
|
| 58 |
+
{%- set tool_call = tool_call.function %}
|
| 59 |
+
{%- endif %}
|
| 60 |
+
{{- '<tool_call>\n{"name": "' }}
|
| 61 |
+
{{- tool_call.name }}
|
| 62 |
+
{{- '", "arguments": ' }}
|
| 63 |
+
{%- if tool_call.arguments is string %}
|
| 64 |
+
{{- tool_call.arguments }}
|
| 65 |
+
{%- else %}
|
| 66 |
+
{{- tool_call.arguments | tojson }}
|
| 67 |
+
{%- endif %}
|
| 68 |
+
{{- '}\n</tool_call>' }}
|
| 69 |
+
{%- endfor %}
|
| 70 |
+
{%- endif %}
|
| 71 |
+
{{- '<|im_end|>\n' }}
|
| 72 |
+
{%- elif message.role == "tool" %}
|
| 73 |
+
{%- if loop.first or (messages[loop.index0 - 1].role != "tool") %}
|
| 74 |
+
{{- '<|im_start|>user' }}
|
| 75 |
+
{%- endif %}
|
| 76 |
+
{{- '\n<tool_response>\n' }}
|
| 77 |
+
{{- content }}
|
| 78 |
+
{{- '\n</tool_response>' }}
|
| 79 |
+
{%- if loop.last or (messages[loop.index0 + 1].role != "tool") %}
|
| 80 |
+
{{- '<|im_end|>\n' }}
|
| 81 |
+
{%- endif %}
|
| 82 |
+
{%- endif %}
|
| 83 |
+
{%- endfor %}
|
| 84 |
+
{%- if add_generation_prompt %}
|
| 85 |
+
{{- '<|im_start|>assistant\n' }}
|
| 86 |
+
{%- if enable_thinking is defined and enable_thinking is false %}
|
| 87 |
+
{{- '<think>\n\n</think>\n\n' }}
|
| 88 |
+
{%- endif %}
|
| 89 |
+
{%- endif %}
|
tokenizer/tokenizer.json
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:be75606093db2094d7cd20f3c2f385c212750648bd6ea4fb2bf507a6a4c55506
|
| 3 |
+
size 11422650
|
tokenizer/tokenizer_config.json
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"add_prefix_space": false,
|
| 3 |
+
"backend": "tokenizers",
|
| 4 |
+
"bos_token": null,
|
| 5 |
+
"clean_up_tokenization_spaces": false,
|
| 6 |
+
"eos_token": "<|im_end|>",
|
| 7 |
+
"errors": "replace",
|
| 8 |
+
"extra_special_tokens": [
|
| 9 |
+
"<|im_start|>",
|
| 10 |
+
"<|im_end|>",
|
| 11 |
+
"<|object_ref_start|>",
|
| 12 |
+
"<|object_ref_end|>",
|
| 13 |
+
"<|box_start|>",
|
| 14 |
+
"<|box_end|>",
|
| 15 |
+
"<|quad_start|>",
|
| 16 |
+
"<|quad_end|>",
|
| 17 |
+
"<|vision_start|>",
|
| 18 |
+
"<|vision_end|>",
|
| 19 |
+
"<|vision_pad|>",
|
| 20 |
+
"<|image_pad|>",
|
| 21 |
+
"<|video_pad|>"
|
| 22 |
+
],
|
| 23 |
+
"is_local": false,
|
| 24 |
+
"model_max_length": 131072,
|
| 25 |
+
"pad_token": "<|endoftext|>",
|
| 26 |
+
"split_special_tokens": false,
|
| 27 |
+
"tokenizer_class": "Qwen2Tokenizer",
|
| 28 |
+
"unk_token": null
|
| 29 |
+
}
|
transformer/config.json
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_class_name": "CosmosTransformer3DModel",
|
| 3 |
+
"_diffusers_version": "0.37.0",
|
| 4 |
+
"in_channels": 16,
|
| 5 |
+
"out_channels": 16,
|
| 6 |
+
"num_attention_heads": 16,
|
| 7 |
+
"attention_head_dim": 128,
|
| 8 |
+
"num_layers": 28,
|
| 9 |
+
"mlp_ratio": 4.0,
|
| 10 |
+
"text_embed_dim": 1024,
|
| 11 |
+
"adaln_lora_dim": 256,
|
| 12 |
+
"max_size": [
|
| 13 |
+
128,
|
| 14 |
+
240,
|
| 15 |
+
240
|
| 16 |
+
],
|
| 17 |
+
"patch_size": [
|
| 18 |
+
1,
|
| 19 |
+
2,
|
| 20 |
+
2
|
| 21 |
+
],
|
| 22 |
+
"rope_scale": [
|
| 23 |
+
1.0,
|
| 24 |
+
4.0,
|
| 25 |
+
4.0
|
| 26 |
+
],
|
| 27 |
+
"concat_padding_mask": true,
|
| 28 |
+
"extra_pos_embed_type": null
|
| 29 |
+
}
|
transformer/diffusion_pytorch_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:e9c0b348c119e44dcc26589102ad5ca64d26ac84d5db3b743d29f0fa2fc2f8b2
|
| 3 |
+
size 3912877072
|
vae/config.json
ADDED
|
@@ -0,0 +1,56 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"_class_name": "AutoencoderKLWan",
|
| 3 |
+
"_diffusers_version": "0.33.0.dev0",
|
| 4 |
+
"attn_scales": [],
|
| 5 |
+
"base_dim": 96,
|
| 6 |
+
"dim_mult": [
|
| 7 |
+
1,
|
| 8 |
+
2,
|
| 9 |
+
4,
|
| 10 |
+
4
|
| 11 |
+
],
|
| 12 |
+
"dropout": 0.0,
|
| 13 |
+
"latents_mean": [
|
| 14 |
+
-0.7571,
|
| 15 |
+
-0.7089,
|
| 16 |
+
-0.9113,
|
| 17 |
+
0.1075,
|
| 18 |
+
-0.1745,
|
| 19 |
+
0.9653,
|
| 20 |
+
-0.1517,
|
| 21 |
+
1.5508,
|
| 22 |
+
0.4134,
|
| 23 |
+
-0.0715,
|
| 24 |
+
0.5517,
|
| 25 |
+
-0.3632,
|
| 26 |
+
-0.1922,
|
| 27 |
+
-0.9497,
|
| 28 |
+
0.2503,
|
| 29 |
+
-0.2921
|
| 30 |
+
],
|
| 31 |
+
"latents_std": [
|
| 32 |
+
2.8184,
|
| 33 |
+
1.4541,
|
| 34 |
+
2.3275,
|
| 35 |
+
2.6558,
|
| 36 |
+
1.2196,
|
| 37 |
+
1.7708,
|
| 38 |
+
2.6052,
|
| 39 |
+
2.0743,
|
| 40 |
+
3.2687,
|
| 41 |
+
2.1526,
|
| 42 |
+
2.8652,
|
| 43 |
+
1.5579,
|
| 44 |
+
1.6382,
|
| 45 |
+
1.1253,
|
| 46 |
+
2.8251,
|
| 47 |
+
1.916
|
| 48 |
+
],
|
| 49 |
+
"num_res_blocks": 2,
|
| 50 |
+
"temperal_downsample": [
|
| 51 |
+
false,
|
| 52 |
+
true,
|
| 53 |
+
true
|
| 54 |
+
],
|
| 55 |
+
"z_dim": 16
|
| 56 |
+
}
|
vae/diffusion_pytorch_model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:3b5bf326a6c4f66fb2b2250687fdccd1f126ee7c977d2f0170cb56fdacc70a9a
|
| 3 |
+
size 253806934
|