File size: 15,859 Bytes
a48bc74 | 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 | """
The Ideogram 4 transformer is a NextDiT/Lumina2-family single-stream model
consumes Qwen3-VL hidden-state features (concatenated from 13 layers -> 53248 dims)
packs ``[text tokens, image tokens]`` into one sequence with block-diagonal segment attention and 3D interleaved MRoPE.
"""
from __future__ import annotations
import math
import torch
import torch.nn as nn
import torch.nn.functional as F
import comfy.model_management
import comfy.ops
import comfy.patcher_extension
import comfy.quant_ops
from comfy.ldm.lumina.model import FeedForward
from comfy.ldm.modules.attention import optimized_attention_masked
from comfy.text_encoders.llama import precompute_freqs_cis
# Per-token role indicators
SEQUENCE_PADDING_INDICATOR = -1
OUTPUT_IMAGE_INDICATOR = 2
LLM_TOKEN_INDICATOR = 3
# Image grid coordinates are offset so they never collide with text positions
IMAGE_POSITION_OFFSET = 65536
def _split_half_rope_matrix(freqs_cis):
cos, sin, neg_sin = freqs_cis
half_dim = sin.shape[-1]
matrix = torch.stack(
(cos[..., :half_dim], neg_sin, sin, cos[..., half_dim:]), dim=-1
)
return matrix.reshape(*matrix.shape[:-1], 2, 2).unsqueeze(2)
def _apply_rope_split_half1(x, freqs_cis):
x_dtype = x.dtype
x = x.reshape(*x.shape[:-1], 2, -1).movedim(-2, -1).unsqueeze(-2).to(freqs_cis.dtype)
output = freqs_cis[..., 0] * x[..., 0] + freqs_cis[..., 1] * x[..., 1]
return output.movedim(-1, -2).reshape(*x.shape[:-3], -1).to(x_dtype)
class Ideogram4Attention(nn.Module):
def __init__(self, hidden_size, num_heads, eps=1e-5, dtype=None, device=None, operations=None):
super().__init__()
self.num_heads = num_heads
self.head_dim = hidden_size // num_heads
self.hidden_size = hidden_size
self.qkv = operations.Linear(hidden_size, hidden_size * 3, bias=False, dtype=dtype, device=device)
self.norm_q = operations.RMSNorm(self.head_dim, eps=eps, elementwise_affine=True, dtype=dtype, device=device)
self.norm_k = operations.RMSNorm(self.head_dim, eps=eps, elementwise_affine=True, dtype=dtype, device=device)
self.o = operations.Linear(hidden_size, hidden_size, bias=False, dtype=dtype, device=device)
def forward(self, x, attn_mask, freqs_cis, transformer_options={}):
batch_size, seq_len, _ = x.shape
qkv = self.qkv(x).view(batch_size, seq_len, 3, self.num_heads, self.head_dim)
q, k, v = qkv.unbind(dim=2)
if comfy.model_management.in_training:
q = _apply_rope_split_half1(self.norm_q(q), freqs_cis)
k = _apply_rope_split_half1(self.norm_k(k), freqs_cis)
else:
q_scale, _, q_offload_stream = comfy.ops.cast_bias_weight(self.norm_q, q, offloadable=True)
k_scale, _, k_offload_stream = comfy.ops.cast_bias_weight(self.norm_k, k, offloadable=True)
q, k = comfy.quant_ops.ck.rms_rope_split_half(
q, k, freqs_cis, q_scale, k_scale, self.norm_q.eps
)
comfy.ops.uncast_bias_weight(self.norm_q, q_scale, None, q_offload_stream)
comfy.ops.uncast_bias_weight(self.norm_k, k_scale, None, k_offload_stream)
# (B, heads, L, head_dim)
q = q.transpose(1, 2)
k = k.transpose(1, 2)
v = v.transpose(1, 2)
out = optimized_attention_masked(q, k, v, self.num_heads, attn_mask, skip_reshape=True, transformer_options=transformer_options)
return self.o(out)
class Ideogram4TransformerBlock(nn.Module):
def __init__(self, hidden_size, intermediate_size, num_heads, norm_eps, adaln_dim, dtype=None, device=None, operations=None):
super().__init__()
self.attention = Ideogram4Attention(hidden_size, num_heads, eps=1e-5, dtype=dtype, device=device, operations=operations)
self.feed_forward = FeedForward(
dim=hidden_size, hidden_dim=intermediate_size, multiple_of=1, ffn_dim_multiplier=None,
operation_settings={"operations": operations, "dtype": dtype, "device": device},
)
self.attention_norm1 = operations.RMSNorm(hidden_size, eps=norm_eps, elementwise_affine=True, dtype=dtype, device=device)
self.ffn_norm1 = operations.RMSNorm(hidden_size, eps=norm_eps, elementwise_affine=True, dtype=dtype, device=device)
self.attention_norm2 = operations.RMSNorm(hidden_size, eps=norm_eps, elementwise_affine=True, dtype=dtype, device=device)
self.ffn_norm2 = operations.RMSNorm(hidden_size, eps=norm_eps, elementwise_affine=True, dtype=dtype, device=device)
self.adaln_modulation = operations.Linear(adaln_dim, 4 * hidden_size, bias=True, dtype=dtype, device=device)
def forward(self, x, attn_mask, freqs_cis, adaln_input, transformer_options={}):
mod = self.adaln_modulation(adaln_input)
scale_msa, gate_msa, scale_mlp, gate_mlp = mod.chunk(4, dim=-1)
gate_msa = torch.tanh(gate_msa)
gate_mlp = torch.tanh(gate_mlp)
scale_msa = 1.0 + scale_msa
scale_mlp = 1.0 + scale_mlp
attn_out = self.attention(self.attention_norm1(x) * scale_msa, attn_mask, freqs_cis, transformer_options=transformer_options)
x = x + gate_msa * self.attention_norm2(attn_out)
x = x + gate_mlp * self.ffn_norm2(self.feed_forward(self.ffn_norm1(x) * scale_mlp))
return x
def _sinusoidal_embedding(t, dim, scale=1e4):
t = t.to(torch.float32)
half = dim // 2
freq = math.log(scale) / (half - 1)
freq = torch.exp(torch.arange(half, dtype=torch.float32, device=t.device) * -freq)
emb = t.unsqueeze(-1) * freq
emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=-1)
if dim % 2 == 1:
emb = F.pad(emb, (0, 1))
return emb
class Ideogram4EmbedScalar(nn.Module):
def __init__(self, dim, input_range=(0.0, 1.0), dtype=None, device=None, operations=None):
super().__init__()
self.dim = dim
self.range_min, self.range_max = input_range
self.mlp_in = operations.Linear(dim, dim, bias=True, dtype=dtype, device=device)
self.mlp_out = operations.Linear(dim, dim, bias=True, dtype=dtype, device=device)
def forward(self, x, dtype):
x = x.to(torch.float32)
scaled = 1e4 * (x - self.range_min) / (self.range_max - self.range_min)
emb = _sinusoidal_embedding(scaled, self.dim)
emb = emb.to(dtype)
emb = F.silu(self.mlp_in(emb))
return self.mlp_out(emb)
class Ideogram4FinalLayer(nn.Module):
def __init__(self, hidden_size, out_channels, adaln_dim, dtype=None, device=None, operations=None):
super().__init__()
self.norm_final = operations.LayerNorm(hidden_size, eps=1e-6, elementwise_affine=False, dtype=dtype, device=device)
self.linear = operations.Linear(hidden_size, out_channels, bias=True, dtype=dtype, device=device)
self.adaln_modulation = operations.Linear(adaln_dim, hidden_size, bias=True, dtype=dtype, device=device)
def forward(self, x, c):
scale = 1.0 + self.adaln_modulation(F.silu(c))
return self.linear(self.norm_final(x) * scale)
class Ideogram4Transformer(nn.Module):
"""A single Ideogram 4 backbone operating on a packed token sequence."""
def __init__(self, emb_dim, num_layers, num_heads, intermediate_size, adaln_dim,
in_channels, llm_features_dim, rope_theta, mrope_section, norm_eps,
dtype=None, device=None, operations=None):
super().__init__()
self.head_dim = emb_dim // num_heads
self.rope_theta = rope_theta
self.mrope_section = tuple(mrope_section)
self.input_proj = operations.Linear(in_channels, emb_dim, bias=True, dtype=dtype, device=device)
self.llm_cond_norm = operations.RMSNorm(llm_features_dim, eps=1e-6, elementwise_affine=True, dtype=dtype, device=device)
self.llm_cond_proj = operations.Linear(llm_features_dim, emb_dim, bias=True, dtype=dtype, device=device)
self.t_embedding = Ideogram4EmbedScalar(emb_dim, input_range=(0.0, 1.0), dtype=dtype, device=device, operations=operations)
self.adaln_proj = operations.Linear(emb_dim, adaln_dim, bias=True, dtype=dtype, device=device)
self.embed_image_indicator = operations.Embedding(2, emb_dim, dtype=dtype, device=device)
self.layers = nn.ModuleList([
Ideogram4TransformerBlock(emb_dim, intermediate_size, num_heads, norm_eps, adaln_dim,
dtype=dtype, device=device, operations=operations)
for _ in range(num_layers)
])
self.final_layer = Ideogram4FinalLayer(emb_dim, in_channels, adaln_dim, dtype=dtype, device=device, operations=operations)
def _backbone(self, llm_features, x, t, position_ids, attn_mask, indicator, transformer_options={}):
indicator = indicator.to(torch.long)
output_image_mask = (indicator == OUTPUT_IMAGE_INDICATOR).to(x.dtype).unsqueeze(-1)
x = x * output_image_mask
h = self.input_proj(x) * output_image_mask
t_cond = self.t_embedding(t, dtype=x.dtype)
if t.dim() == 1:
t_cond = t_cond.unsqueeze(1)
adaln_input = F.silu(self.adaln_proj(t_cond))
# h is zero on the text rows (content lives only on image rows), add writes the text features in place
if llm_features is not None:
L_text = llm_features.shape[1]
text_mask = (indicator[:, :L_text] == LLM_TOKEN_INDICATOR).to(x.dtype).unsqueeze(-1)
llm = self.llm_cond_norm(llm_features * text_mask)
llm = self.llm_cond_proj(llm) * text_mask
h[:, :L_text] = h[:, :L_text] + llm
h = h + self.embed_image_indicator((indicator == OUTPUT_IMAGE_INDICATOR).to(torch.long), out_dtype=h.dtype)
# Qwen3-VL interleaved MRoPE; position_ids (B, L, 3) -> (3, L) (same across batch).
freqs_cis = precompute_freqs_cis(
self.head_dim, position_ids[0].transpose(0, 1), self.rope_theta,
rope_dims=self.mrope_section, interleaved_mrope=True, device=position_ids.device,
)
freqs_cis = _split_half_rope_matrix(freqs_cis)
if attn_mask is not None and attn_mask.dtype == torch.bool:
attn_mask = torch.zeros_like(attn_mask, dtype=h.dtype).masked_fill_(~attn_mask, -torch.finfo(h.dtype).max)
for layer in self.layers:
h = layer(h, attn_mask, freqs_cis, adaln_input, transformer_options=transformer_options)
return self.final_layer(h, adaln_input)
class Ideogram4Transformer2DModel(Ideogram4Transformer):
"""Ideogram 4 single-stream DiT.
Runs a packed ``[text, image]`` sequence when text context is supplied, or an image-only sequence when ``context is None``.
"""
def __init__(self, image_model=None, in_channels=128, num_layers=34, num_attention_heads=18, attention_head_dim=256, intermediate_size=12288,
adaln_dim=512, llm_features_dim=53248, rope_theta=5000000, mrope_section=(24, 20, 20), norm_eps=1e-5,
dtype=None, device=None, operations=None, **kwargs):
emb_dim = num_attention_heads * attention_head_dim
super().__init__(
emb_dim=emb_dim, num_layers=num_layers, num_heads=num_attention_heads,
intermediate_size=intermediate_size, adaln_dim=adaln_dim, in_channels=in_channels,
llm_features_dim=llm_features_dim, rope_theta=rope_theta, mrope_section=mrope_section,
norm_eps=norm_eps, dtype=dtype, device=device, operations=operations)
self.dtype = dtype
self.in_channels = in_channels
self.out_channels = in_channels
# 128-dim token = patch (2x2) * ae_channels (32).
self.patch_size = 2
self.ae_channels = in_channels // (self.patch_size * self.patch_size)
def _img_to_tokens(self, x):
B, C, gh, gw = x.shape
x = x.view(B, self.ae_channels, self.patch_size, self.patch_size, gh, gw)
x = x.permute(0, 4, 5, 2, 3, 1) # (B, gh, gw, pi, pj, c)
return x.reshape(B, gh * gw, C)
def _tokens_to_img(self, tokens, gh, gw):
B = tokens.shape[0]
C = tokens.shape[-1]
x = tokens.reshape(B, gh, gw, self.patch_size, self.patch_size, self.ae_channels)
x = x.permute(0, 5, 3, 4, 1, 2) # (B, c, pi, pj, gh, gw)
return x.reshape(B, C, gh, gw)
def _image_position_ids(self, gh, gw, device):
h_idx = torch.arange(gh, device=device).view(-1, 1).expand(gh, gw).reshape(-1)
w_idx = torch.arange(gw, device=device).view(1, -1).expand(gh, gw).reshape(-1)
t_idx = torch.zeros_like(h_idx)
return torch.stack([t_idx, h_idx, w_idx], dim=1) + IMAGE_POSITION_OFFSET # (L_img, 3)
def _run_conditional(self, x_chunk, context_chunk, attn_mask_chunk, t_chunk, gh, gw, transformer_options):
B = x_chunk.shape[0]
device = x_chunk.device
img_tokens = self._img_to_tokens(x_chunk)
L_img = img_tokens.shape[1]
L_text = context_chunk.shape[1]
L = L_text + L_img
latent_dim = img_tokens.shape[-1]
x_full = torch.zeros(B, L, latent_dim, dtype=img_tokens.dtype, device=device)
x_full[:, L_text:] = img_tokens
text_pos = torch.arange(L_text, device=device).view(-1, 1).expand(L_text, 3)
img_pos = self._image_position_ids(gh, gw, device)
position_ids = torch.cat([text_pos, img_pos], dim=0).unsqueeze(0).expand(B, L, 3)
indicator = torch.empty(B, L, dtype=torch.long, device=device)
indicator[:, :L_text] = LLM_TOKEN_INDICATOR
indicator[:, L_text:] = OUTPUT_IMAGE_INDICATOR
attn_mask = None
if attn_mask_chunk is not None:
segment_ids = torch.ones(B, L, dtype=torch.long, device=device)
pad = (attn_mask_chunk == 0)
segment_ids[:, :L_text][pad] = SEQUENCE_PADDING_INDICATOR
indicator[:, :L_text][pad] = 0
# Block-diagonal mask from segment ids: (B, 1, L, L), True = attend.
attn_mask = (segment_ids.unsqueeze(2) == segment_ids.unsqueeze(1)).unsqueeze(1)
out = self._backbone(context_chunk, x_full, t_chunk, position_ids, attn_mask, indicator,
transformer_options=transformer_options)
return self._tokens_to_img(out[:, L_text:], gh, gw)
def _run_image_only(self, x_chunk, t_chunk, gh, gw, transformer_options):
B = x_chunk.shape[0]
device = x_chunk.device
img_tokens = self._img_to_tokens(x_chunk)
L_img = img_tokens.shape[1]
position_ids = self._image_position_ids(gh, gw, device).unsqueeze(0).expand(B, L_img, 3)
indicator = torch.full((B, L_img), OUTPUT_IMAGE_INDICATOR, dtype=torch.long, device=device)
# Image-only sequence is a single segment -> no mask, full attention, no LLM context.
out = self._backbone(None, img_tokens, t_chunk, position_ids, None, indicator, transformer_options=transformer_options)
return self._tokens_to_img(out, gh, gw)
def forward(self, x, timesteps, context=None, attention_mask=None, transformer_options={}, **kwargs):
return comfy.patcher_extension.WrapperExecutor.new_class_executor(
self._forward,
self,
comfy.patcher_extension.get_all_wrappers(comfy.patcher_extension.WrappersMP.DIFFUSION_MODEL, transformer_options),
).execute(x, timesteps, context, attention_mask, transformer_options, **kwargs)
def _forward(self, x, timesteps, context=None, attention_mask=None, transformer_options={}, **kwargs):
bs, c, gh, gw = x.shape
timesteps = 1.0 - timesteps
# unconditional pass
if context is None:
return -self._run_image_only(x, timesteps, gh, gw, transformer_options)
return -self._run_conditional(x, context, attention_mask, timesteps, gh, gw, transformer_options)
|