File size: 12,532 Bytes
e1cb4af 325ed02 e1cb4af 325ed02 e1cb4af 325ed02 e1cb4af 325ed02 e1cb4af 325ed02 e1cb4af 325ed02 e1cb4af 4d84cbc e1cb4af 9de6e96 4d84cbc e1cb4af 325ed02 e1cb4af 6e55001 325ed02 e1cb4af 325ed02 |
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 |
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.utils.checkpoint
from torch.nn import LayerNorm
from transformers.modeling_utils import PreTrainedModel
from .configuration_dots import DotsVisionConfig
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_rotary_pos_emb_vision(tensor: torch.Tensor, freqs: torch.Tensor) -> torch.Tensor:
orig_dtype = tensor.dtype
tensor = tensor.float()
cos = freqs.cos()
sin = freqs.sin()
cos = cos.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
sin = sin.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
output = (tensor * cos) + (rotate_half(tensor) * sin)
output = output.to(orig_dtype)
return output
class VisionRotaryEmbedding(nn.Module):
def __init__(self, dim: int, theta: float = 10000.0) -> None:
super().__init__()
inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
self.register_buffer("inv_freq", inv_freq, persistent=False)
def forward(self, seqlen: int) -> torch.Tensor:
seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
freqs = torch.outer(seq, self.inv_freq)
return freqs
class PatchMerger(nn.Module):
def __init__(
self,
dim: int,
context_dim: int,
spatial_merge_size: int = 2,
pre_norm="layernorm",
init_merger_std=None,
) -> None:
super().__init__()
self.hidden_size = context_dim * (spatial_merge_size ** 2)
self.pre_norm = pre_norm
if self.pre_norm == "layernorm":
self.ln_q = LayerNorm(context_dim, eps=1e-6)
elif self.pre_norm == "rmsnorm":
self.ln_q = RMSNorm(context_dim, eps=1e-6)
else:
print("no norm in patch merger")
self.mlp = nn.Sequential(
nn.Linear(self.hidden_size, self.hidden_size),
nn.GELU(),
nn.Linear(self.hidden_size, dim),
)
if init_merger_std is not None:
nn.init.normal_(self.mlp[0].weight, mean=0.0, std=init_merger_std)
nn.init.zeros_(self.mlp[0].bias)
nn.init.normal_(self.mlp[2].weight, mean=0.0, std=init_merger_std)
nn.init.zeros_(self.mlp[2].bias)
def forward(self, x: torch.Tensor) -> torch.Tensor:
if self.pre_norm:
x = self.mlp(self.ln_q(x).view(-1, self.hidden_size))
else:
x = self.mlp(x.view(-1, self.hidden_size))
return x
class VisionSdpaAttention(nn.Module):
def __init__(self, config, dim: int, num_heads: int = 16, bias=True) -> None:
super().__init__()
self.num_heads = num_heads
self.qkv = nn.Linear(dim, dim * 3, bias=bias)
self.proj = nn.Linear(dim, dim, bias=bias)
self.config = config
def forward(
self,
hidden_states: torch.Tensor,
cu_seqlens: torch.Tensor,
rotary_pos_emb: torch.Tensor = None,
) -> torch.Tensor:
seq_length = hidden_states.shape[0]
q, k, v = self.qkv(hidden_states).reshape(seq_length, 3, self.num_heads, -1).permute(1, 0, 2, 3).unbind(0)
q = apply_rotary_pos_emb_vision(q.unsqueeze(0), rotary_pos_emb).squeeze(0)
k = apply_rotary_pos_emb_vision(k.unsqueeze(0), rotary_pos_emb).squeeze(0)
attention_mask = torch.zeros([1, seq_length, seq_length], device=q.device, dtype=torch.bool)
for i in range(1, len(cu_seqlens)):
attention_mask[..., cu_seqlens[i - 1]: cu_seqlens[i], cu_seqlens[i - 1]: cu_seqlens[i]] = True
q = q.transpose(0, 1).unsqueeze(0) # (1, num_heads, seq_length, head_dim)
k = k.transpose(0, 1).unsqueeze(0)
v = v.transpose(0, 1).unsqueeze(0)
attn_output = F.scaled_dot_product_attention(q, k, v, attention_mask, dropout_p=0.0)
attn_output = attn_output.squeeze(0).transpose(0, 1) # (seq_length, num_heads, head_dim)
attn_output = attn_output.reshape(seq_length, -1)
attn_output = self.proj(attn_output)
return attn_output
DOTS_VISION_ATTENTION_CLASSES = {
"sdpa": VisionSdpaAttention,
}
class RMSNorm(nn.Module):
def __init__(self, dim: int, eps: float = 1e-6):
super().__init__()
self.weight = nn.Parameter(torch.ones(dim))
self.eps = eps
def forward(self, x: torch.Tensor) -> torch.Tensor:
output = self._norm(x.float()).type_as(x)
return output * self.weight
def extra_repr(self) -> str:
return f"{tuple(self.weight.shape)}, eps={self.eps}"
def _norm(self, x: torch.Tensor) -> torch.Tensor:
return x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True) + self.eps)
class DotsSwiGLUFFN(nn.Module):
def __init__(self, config):
super().__init__()
hidden_features = config.intermediate_size
in_features = config.embed_dim
bias = config.use_bias
self.fc1 = nn.Linear(in_features, hidden_features, bias=bias)
self.fc2 = nn.Linear(hidden_features, in_features, bias=bias)
self.fc3 = nn.Linear(in_features, hidden_features, bias=bias)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = F.silu(self.fc1(x)) * self.fc3(x)
x = self.fc2(x)
return x
class DotsPatchEmbed(nn.Module):
def __init__(self, config):
super().__init__()
self.num_channels = config.num_channels
self.patch_size = config.patch_size
self.temporal_patch_size = config.temporal_patch_size
self.embed_dim = config.embed_dim
self.config = config
self.proj = nn.Conv2d(
config.num_channels,
config.embed_dim,
kernel_size=(config.patch_size, config.patch_size),
stride=(config.patch_size, config.patch_size),
)
self.norm = RMSNorm(config.embed_dim, eps=config.rms_norm_eps)
def forward(self, x: torch.Tensor, grid_thw=None) -> torch.Tensor:
x = x.view(-1, self.num_channels, self.temporal_patch_size, self.patch_size, self.patch_size)[:, :, 0]
x = self.proj(x).view(-1, self.embed_dim)
x = self.norm(x)
return x
class DotsViTPreprocessor(nn.Module):
def __init__(self, config):
super().__init__()
self.patch_h = config.patch_size
self.patch_w = config.patch_size
self.embed_dim = config.embed_dim
self.config = config
self.patchifier = DotsPatchEmbed(config)
def forward(self, x: torch.Tensor, grid_thw=None) -> torch.Tensor:
tokens = self.patchifier(x, grid_thw)
return tokens
class DotsVisionBlock(nn.Module):
def __init__(self, config, attn_implementation: str = "flash_attention_2"):
super().__init__()
attn_implementation = "sdpa" # hardcode sdpa
if attn_implementation == "flash_attention_2" and not flash_attn_available:
# fallback to eager
attn_implementation = "eager"
print("flash attention not available! fallback to eager implementation ")
if attn_implementation == "ascend_fa" and not npu_available:
attn_implementation = "eager"
print("flash attention not available! fallback to eager implementation ")
self.attn = DOTS_VISION_ATTENTION_CLASSES[attn_implementation](
config, config.embed_dim, num_heads=config.num_attention_heads, bias=config.use_bias
)
self.norm1 = RMSNorm(config.embed_dim, eps=config.rms_norm_eps)
self.mlp = DotsSwiGLUFFN(config)
self.norm2 = RMSNorm(config.embed_dim, eps=config.rms_norm_eps)
def forward(self, hidden_states, cu_seqlens, rotary_pos_emb) -> torch.Tensor:
hidden_states = hidden_states + self.attn(
self.norm1(hidden_states), cu_seqlens=cu_seqlens, rotary_pos_emb=rotary_pos_emb
)
hidden_states = hidden_states + self.mlp(self.norm2(hidden_states))
return hidden_states
class DotsVisionTransformer(PreTrainedModel):
def __init__(self, config: DotsVisionConfig) -> None:
super().__init__(config)
self.config = config
self.spatial_merge_size = config.spatial_merge_size
self.patch_embed = DotsViTPreprocessor(config)
self._init_weights(self.patch_embed.patchifier.proj)
head_dim = config.embed_dim // config.num_attention_heads
self.rotary_pos_emb = VisionRotaryEmbedding(head_dim // 2)
_num_hidden_layers = config.num_hidden_layers
self.blocks = nn.ModuleList(
[DotsVisionBlock(config, config.attn_implementation) for _ in range(_num_hidden_layers)]
)
if self.config.post_norm:
self.post_trunk_norm = RMSNorm(config.embed_dim, eps=config.rms_norm_eps)
self.merger = PatchMerger(
dim=config.hidden_size,
context_dim=config.embed_dim,
spatial_merge_size=config.spatial_merge_size,
init_merger_std=self.config.init_merger_std,
)
self.gradient_checkpointing = False
self._gradient_checkpointing_func = torch.utils.checkpoint.checkpoint
def _init_weights(self, module):
std = self.config.initializer_range
if isinstance(module, (nn.Linear, nn.Conv3d)):
module.weight.data.normal_(mean=0.0, std=std)
if module.bias is not None:
module.bias.data.zero_()
elif isinstance(module, nn.Embedding):
module.weight.data.normal_(mean=0.0, std=std)
if module.padding_idx is not None:
module.weight.data[module.padding_idx].zero_()
@property
def dtype(self) -> torch.dtype:
return self.blocks[0].mlp.fc2.weight.dtype
@property
def device(self) -> torch.device:
return self.blocks[0].mlp.fc2.weight.device
def get_pos_ids_by_grid(self, grid_thw):
pos_ids = []
for t, h, w in grid_thw:
hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
hpos_ids = hpos_ids.reshape(
h // self.spatial_merge_size,
self.spatial_merge_size,
w // self.spatial_merge_size,
self.spatial_merge_size,
)
hpos_ids = hpos_ids.permute(0, 2, 1, 3)
hpos_ids = hpos_ids.flatten()
wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
wpos_ids = wpos_ids.reshape(
h // self.spatial_merge_size,
self.spatial_merge_size,
w // self.spatial_merge_size,
self.spatial_merge_size,
)
wpos_ids = wpos_ids.permute(0, 2, 1, 3)
wpos_ids = wpos_ids.flatten()
pos_ids.append(
torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1)
)
return pos_ids
def rot_pos_emb(self, grid_thw):
pos_ids = self.get_pos_ids_by_grid(grid_thw)
pos_ids = torch.cat(pos_ids, dim=0)
max_grid_size = grid_thw[:, 1:].max()
rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
return rotary_pos_emb
def forward(self, hidden_states: torch.Tensor, grid_thw: torch.Tensor, bf16=True) -> torch.Tensor:
if bf16:
hidden_states = hidden_states.bfloat16()
hidden_states = self.patch_embed(hidden_states, grid_thw)
rotary_pos_emb = self.rot_pos_emb(grid_thw)
cu_seqlens = torch.repeat_interleave(grid_thw[:, 1] * grid_thw[:, 2], grid_thw[:, 0]).cumsum(
dim=0,
dtype=grid_thw.dtype if torch.jit.is_tracing() else torch.int32,
)
cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
for blk in self.blocks:
if self.gradient_checkpointing and self.training:
hidden_states = self._gradient_checkpointing_func(
blk.__call__,
hidden_states,
cu_seqlens,
rotary_pos_emb,
)
else:
hidden_states = blk(hidden_states, cu_seqlens=cu_seqlens, rotary_pos_emb=rotary_pos_emb)
if self.config.post_norm:
hidden_states = self.post_trunk_norm(hidden_states)
hidden_states = self.merger(hidden_states)
return hidden_states
|