Spaces:
Running on Zero
Running on Zero
File size: 14,933 Bytes
834d1cf | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 | import torch
import torch.nn as nn
from einops import rearrange
from timm.models.vision_transformer import PatchEmbed
def rotate_half(x):
"""Rotates half the hidden dims of the input."""
x1 = x[..., : x.shape[-1] // 2]
x2 = x[..., x.shape[-1] // 2 :]
return torch.cat((-x2, x1), dim=-1)
def apply_multimodal_rotary_pos_emb(
q: torch.Tensor,
k: torch.Tensor,
cos: torch.Tensor,
sin: torch.Tensor,
mrope_section: list[int],
unsqueeze_dim: int = 2
) -> tuple[torch.Tensor, torch.Tensor]:
"""
[重写] 严格按照 Qwen2.5-VL 的实现。
将 3D RoPE (T, H, W) 应用于 query 和 key。
Args:
q (`torch.Tensor`): query (B, N_img, H, Hc)
k (`torch.Tensor`): key (B, N_img, H, Hc)
cos (`torch.Tensor`): cosine (3, B, N_img, Hc)
sin (`torch.Tensor`): sine (3, B, N_img, Hc)
mrope_section (`List[int]`):
T, H, W 的通道维度列表, e.g., [16, 24, 24].
注意: 2 * sum(mrope_section) 必须等于 Hc.
unsqueeze_dim (`int`, *optional*, defaults to 2):
为 (cos, sin) 增加的广播维度。
在我们的 Attention 模块中, q/k 形状为 (B, N_img, H, Hc),
因此我们使用 unsqueeze_dim=2 使 cos/sin 形状变为 (B, N_img, 1, Hc)
以便在 H (头数) 维度上广播。
(Qwen-VL 原始代码默认为 1, 因为它在 (B, H, N_img, Hc) 上操作)
"""
split_sections = mrope_section * 2
cos_chunks = cos.split(split_sections, dim=-1)
sin_chunks = sin.split(split_sections, dim=-1)
cos_emb = torch.cat(
[m[i % 3] for i, m in enumerate(cos_chunks)],
dim=-1
).unsqueeze(unsqueeze_dim)
sin_emb = torch.cat(
[m[i % 3] for i, m in enumerate(sin_chunks)],
dim=-1
).unsqueeze(unsqueeze_dim)
q_embed = (q * cos_emb) + (rotate_half(q) * sin_emb)
k_embed = (k * cos_emb) + (rotate_half(k) * sin_emb)
return q_embed, k_embed
class Multimodal3DRotaryEmbedding(nn.Module):
"""
Qwen2.5-VL 文本模型使用的 3D RoPE 实现。
它接收 3D 坐标 (T, H, W) 并为每个坐标计算独立的 RoPE。
"""
inv_freq: torch.Tensor
def __init__(self, head_dim: int, theta: float = 10000.0, device=None):
"""
dim: 应该是 head_dim (Hc)
"""
super().__init__()
self.head_dim = head_dim
self.theta = theta
inv_freq = 1.0 / (
self.theta ** (torch.arange(0, self.head_dim, 2, dtype=torch.int64).to(device=device, dtype=torch.float) / self.head_dim)
)
self.register_buffer("inv_freq", inv_freq, persistent=False)
self.attention_scaling = 1.0
def forward(self, x: torch.Tensor, position_ids: torch.Tensor):
"""
x: 任意张量，仅用于获取 device 和 dtype
position_ids: [3, B, N_img] (T, H, W 坐标)
"""
inv_freq_expanded = self.inv_freq[None, None, :, None].float().expand(3, position_ids.shape[1], -1, 1)
position_ids_expanded = position_ids.float().unsqueeze(2)
device_type = x.device.type if isinstance(x.device.type, str) and x.device.type != "mps" else "cpu"
with torch.autocast(device_type=device_type, enabled=False):
freqs = (inv_freq_expanded @ position_ids_expanded).transpose(2, 3)
emb = torch.cat((freqs, freqs), dim=-1)
cos = emb.cos() * self.attention_scaling
sin = emb.sin() * self.attention_scaling
return cos.to(dtype=x.dtype), sin.to(dtype=x.dtype)
class RMSNorm(nn.Module):
def __init__(self, hidden_size, eps=1e-6):
super().__init__()
self.weight = nn.Parameter(torch.ones(hidden_size))
self.variance_epsilon = eps
def forward(self, hidden_states):
input_dtype = hidden_states.dtype
hidden_states = hidden_states.to(torch.float32)
variance = hidden_states.pow(2).mean(-1, keepdim=True)
hidden_states = hidden_states * torch.rsqrt(variance + self.variance_epsilon)
return self.weight * hidden_states.to(input_dtype)
class FeedForward(nn.Module):
def __init__(
self,
dim: int,
hidden_dim: int,
):
super().__init__()
self.fc1 = nn.Linear(dim, hidden_dim)
self.act = nn.GELU()
self.fc2 = nn.Linear(hidden_dim, dim)
def forward(self, x):
x = self.fc1(x)
x = self.act(x)
x = self.fc2(x)
return x
class Attention(nn.Module):
def __init__(
self,
dim: int,
num_heads: int = 8,
qkv_bias: bool = False,
qk_norm: bool = True,
attn_drop: float = 0.,
proj_drop: float = 0.,
norm_layer: nn.Module = RMSNorm,
) -> None:
super().__init__()
assert dim % num_heads == 0, "dim should be divisible by num_heads"
self.dim = dim
self.num_heads = num_heads
self.head_dim = dim // num_heads
self.scale = self.head_dim ** -0.5
self.mrope_section = [8,12,12]
self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
self.q_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
self.k_norm = norm_layer(self.head_dim) if qk_norm else nn.Identity()
self.attn_drop = attn_drop
self.proj = nn.Linear(dim, dim)
self.proj_drop = nn.Dropout(proj_drop)
def forward(self, x: torch.Tensor, pos: tuple[torch.Tensor, torch.Tensor], mask=None) -> torch.Tensor:
"""
x: [B, N_total, C] (N_total = N_images + N_query)
pos: (cos, sin)
cos/sin 形状为 [3, B, N_total, Hc]
"""
B, N_total, C = x.shape
qkv = self.qkv(x).reshape(B, N_total, 3, self.num_heads, C // self.num_heads).permute(2, 0, 1, 3, 4)
q, k, v = qkv[0], qkv[1], qkv[2]
q = self.q_norm(q)
k = self.k_norm(k)
cos, sin = pos
q, k = apply_multimodal_rotary_pos_emb(
q,
k,
cos=cos,
sin=sin,
mrope_section=self.mrope_section,
unsqueeze_dim=2
)
q = q.transpose(1, 2)
k = k.transpose(1, 2)
v = v.transpose(1, 2)
x = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=self.attn_drop)
x = x.transpose(1, 2).reshape(B, N_total, C)
x = self.proj(x)
x = self.proj_drop(x)
return x
class Block(nn.Module):
def __init__(self, hidden_size, num_heads, mlp_ratio=4.0):
super().__init__()
self.norm1 = RMSNorm(hidden_size, eps=1e-6)
self.attn = Attention(hidden_size, num_heads=num_heads, qkv_bias=False)
self.norm2 = RMSNorm(hidden_size, eps=1e-6)
self.mlp = FeedForward(hidden_size, int(hidden_size * mlp_ratio))
def forward(self, x, pos, mask=None):
residual = x
x = self.norm1(x)
x = self.attn(x, pos, mask=mask)
x = residual + x
residual = x
x = self.norm2(x)
x = self.mlp(x)
x = residual + x
return x
class QwenImageExtractor(nn.Module):
def __init__(self, hidden_size=1024,num_layers=16,num_heads=16,query_length=256,patch_size=2,in_chans=16,output_dim=3584):
super().__init__()
self.hidden_size = hidden_size
self.num_layers = num_layers
self.num_heads = num_heads
self.head_dim = hidden_size // num_heads
self.output_dim=output_dim
self.patch_embed = PatchEmbed(
img_size = None,
patch_size = patch_size,
in_chans = in_chans,
embed_dim = hidden_size,
strict_img_size=False
)
self.query_length=query_length
self.query = nn.Parameter(torch.randn(1, self.query_length, self.hidden_size))
self.blocks = nn.ModuleList([Block(hidden_size, num_heads) for _ in range(self.num_layers)])
self.norm2 = nn.LayerNorm(hidden_size)
self.output_proj = nn.Linear(hidden_size, output_dim)
self.rotary_pos_emb = Multimodal3DRotaryEmbedding(self.head_dim)
self.cached_pos_ids_shape = None
self.cached_pos_ids = None
def _compute_dynamic_pos_ids(self, H: int, W: int, device: torch.device) -> torch.Tensor:
"""
⭐️ (修改点 4)
根据输入的 H 和 W 动态计算 3D RoPE 的 Position IDs
返回: [3, L_total] (L_images + L_query)
"""
patch_size_h, patch_size_w = self.patch_embed.patch_size
grid_h = H // patch_size_h
grid_w = W // patch_size_w
current_shape = (grid_h, grid_w)
if current_shape == self.cached_pos_ids_shape:
if self.cached_pos_ids.device == device:
return self.cached_pos_ids
else:
self.cached_pos_ids = self.cached_pos_ids.to(device)
return self.cached_pos_ids
num_patches_per_image = grid_h * grid_w
hpos_ids = torch.arange(grid_h, device=device).unsqueeze(1).expand(-1, grid_w).flatten()
wpos_ids = torch.arange(grid_w, device=device).unsqueeze(0).expand(grid_h, -1).flatten()
tpos_ids_0 = torch.zeros(num_patches_per_image, device=device, dtype=torch.long)
pos_ids_0 = torch.stack([tpos_ids_0, hpos_ids, wpos_ids], dim=0)
tpos_ids_1 = torch.ones(num_patches_per_image, device=device, dtype=torch.long)
pos_ids_1 = torch.stack([tpos_ids_1, hpos_ids, wpos_ids], dim=0)
img_pos_ids = torch.cat([pos_ids_0, pos_ids_1], dim=1)
start_idx = max(grid_h, grid_w)
query_indices = torch.arange(start_idx, start_idx + self.query_length, device=device, dtype=torch.long)
query_pos_ids = query_indices.unsqueeze(0).expand(3, -1)
all_pos_ids = torch.cat([img_pos_ids, query_pos_ids], dim=1)
self.cached_pos_ids_shape = current_shape
self.cached_pos_ids = all_pos_ids
return all_pos_ids
def forward(self, x):
"""
x: [B, 2, C, H, W]
"""
B, N, C, H, W = x.shape
assert N == 2, "This model is hardcoded for N=2"
pos_ids = self._compute_dynamic_pos_ids(H, W, device=x.device)
pos_ids_batch = pos_ids.unsqueeze(1).expand(-1, B, -1)
pos_tuple = self.rotary_pos_emb(x, pos_ids_batch)
x = rearrange(x, "B N C H W -> (B N) C H W")
x = self.patch_embed(x)
x = rearrange(x, "(B N) l d -> B (N l) d", B=B, N=N)
x = torch.cat([x, self.query.repeat(B, 1, 1)], dim=1)
if x.shape[1] != pos_ids.shape[1]:
raise ValueError(
f"Total sequence length mismatch. "
f"Input tensor 'x' length is {x.shape[1]}, "
f"but calculated 3D RoPE length is {pos_ids.shape[1]}. "
f"Check H/W ({H}/{W}) vs patch_size ({self.patch_embed.patch_size})."
)
for block in self.blocks:
x = block(x, pos=pos_tuple)
x=x[:, -self.query_length:, :]
x = self.norm2(x)
x = self.output_proj(x)
return x
|