divisor.flux2.model

  1# SPDX-License-Identifier:Apache-2.0
  2# original BFL Flux code from https://github.com/black-forest-labs/flux2
  3
  4from dataclasses import dataclass, field
  5import math
  6
  7from einops import rearrange
  8import torch
  9from torch import Tensor, nn
 10
 11from divisor.layer_dropout import process_blocks_with_dropout
 12
 13
 14@dataclass
 15class Flux2Params:
 16    in_channels: int = 128
 17    context_in_dim: int = 15360
 18    hidden_size: int = 6144
 19    num_heads: int = 48
 20    depth: int = 8
 21    depth_single_blocks: int = 48
 22    axes_dim: list[int] = field(default_factory=lambda: [32, 32, 32, 32])
 23    theta: int = 2000
 24    mlp_ratio: float = 3.0
 25
 26
 27class Flux2(nn.Module):
 28    def __init__(self, params: Flux2Params):
 29        super().__init__()
 30
 31        self.in_channels = params.in_channels
 32        self.out_channels = params.in_channels
 33        if params.hidden_size % params.num_heads != 0:
 34            raise ValueError(f"Hidden size {params.hidden_size} must be divisible by num_heads {params.num_heads}")
 35        pe_dim = params.hidden_size // params.num_heads
 36        if sum(params.axes_dim) != pe_dim:
 37            raise ValueError(f"Got {params.axes_dim} but expected positional dim {pe_dim}")
 38        self.hidden_size = params.hidden_size
 39        self.num_heads = params.num_heads
 40        self.pe_embedder = EmbedND(dim=pe_dim, theta=params.theta, axes_dim=params.axes_dim)
 41        self.img_in = nn.Linear(self.in_channels, self.hidden_size, bias=False)
 42        self.time_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, disable_bias=True)
 43        self.guidance_in = MLPEmbedder(in_dim=256, hidden_dim=self.hidden_size, disable_bias=True)
 44        self.txt_in = nn.Linear(params.context_in_dim, self.hidden_size, bias=False)
 45
 46        self.double_blocks = nn.ModuleList(
 47            [
 48                DoubleStreamBlock(
 49                    self.hidden_size,
 50                    self.num_heads,
 51                    mlp_ratio=params.mlp_ratio,
 52                )
 53                for _ in range(params.depth)
 54            ]
 55        )
 56
 57        self.single_blocks = nn.ModuleList(
 58            [
 59                SingleStreamBlock(
 60                    self.hidden_size,
 61                    self.num_heads,
 62                    mlp_ratio=params.mlp_ratio,
 63                )
 64                for _ in range(params.depth_single_blocks)
 65            ]
 66        )
 67
 68        self.double_stream_modulation_img = Modulation(
 69            self.hidden_size,
 70            double=True,
 71            disable_bias=True,
 72        )
 73        self.double_stream_modulation_txt = Modulation(
 74            self.hidden_size,
 75            double=True,
 76            disable_bias=True,
 77        )
 78        self.single_stream_modulation = Modulation(self.hidden_size, double=False, disable_bias=True)
 79
 80        self.final_layer = LastLayer(
 81            self.hidden_size,
 82            self.out_channels,
 83        )
 84
 85    def forward(
 86        self,
 87        x: Tensor,
 88        x_ids: Tensor,
 89        timesteps: Tensor,
 90        ctx: Tensor,
 91        ctx_ids: Tensor,
 92        guidance: Tensor,
 93        layer_dropouts: list[int] | None = None,
 94    ):
 95        num_txt_tokens = ctx.shape[1]
 96
 97        timestep_emb = timestep_embedding(timesteps, 256)
 98        vec = self.time_in(timestep_emb)
 99        guidance_emb = timestep_embedding(guidance, 256)
100        vec = vec + self.guidance_in(guidance_emb)
101
102        double_block_mod_img = self.double_stream_modulation_img(vec)
103        double_block_mod_txt = self.double_stream_modulation_txt(vec)
104        single_block_mod, _ = self.single_stream_modulation(vec)
105
106        img = self.img_in(x)
107        txt = self.txt_in(ctx)
108
109        pe_x = self.pe_embedder(x_ids)
110        pe_ctx = self.pe_embedder(ctx_ids)
111
112        img, txt = process_blocks_with_dropout(
113            self.double_blocks,
114            layer_dropouts,  # Would need to be added to flux2 forward signature
115            0,
116            "double",
117            lambda block, state: block(state[0], state[1], pe_x, pe_ctx, double_block_mod_img, double_block_mod_txt),
118            (img, txt),
119        )
120
121        img = torch.cat((txt, img), dim=1)
122        pe = torch.cat((pe_ctx, pe_x), dim=2)
123
124        img = process_blocks_with_dropout(
125            self.single_blocks,
126            layer_dropouts,  # Would need to be added to flux2 forward signature
127            len(self.double_blocks),
128            "single",
129            lambda block, state: block(state, pe, single_block_mod),
130            img,
131        )
132        img = img[:, num_txt_tokens:, ...]
133
134        img = self.final_layer(img, vec)
135        return img
136
137
138class SelfAttention(nn.Module):
139    def __init__(
140        self,
141        dim: int,
142        num_heads: int = 8,
143    ):
144        super().__init__()
145        self.num_heads = num_heads
146        head_dim = dim // num_heads
147        self.qkv = nn.Linear(dim, dim * 3, bias=False)
148
149        self.norm = QKNorm(head_dim)
150        self.proj = nn.Linear(dim, dim, bias=False)
151
152
153class SiLUActivation(nn.Module):
154    def __init__(self):
155        super().__init__()
156        self.gate_fn = nn.SiLU()
157
158    def forward(self, x: Tensor) -> Tensor:
159        x1, x2 = x.chunk(2, dim=-1)
160        return self.gate_fn(x1) * x2
161
162
163class Modulation(nn.Module):
164    def __init__(self, dim: int, double: bool, disable_bias: bool = False):
165        super().__init__()
166        self.is_double = double
167        self.multiplier = 6 if double else 3
168        self.lin = nn.Linear(dim, self.multiplier * dim, bias=not disable_bias)
169
170    def forward(self, vec: torch.Tensor):
171        out = self.lin(nn.functional.silu(vec))
172        if out.ndim == 2:
173            out = out[:, None, :]
174        out = out.chunk(self.multiplier, dim=-1)
175        return out[:3], out[3:] if self.is_double else None
176
177
178class LastLayer(nn.Module):
179    def __init__(
180        self,
181        hidden_size: int,
182        out_channels: int,
183    ):
184        super().__init__()
185        self.norm_final = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
186        self.linear = nn.Linear(hidden_size, out_channels, bias=False)
187        self.adaLN_modulation = nn.Sequential(nn.SiLU(), nn.Linear(hidden_size, 2 * hidden_size, bias=False))
188
189    def forward(self, x: torch.Tensor, vec: torch.Tensor) -> torch.Tensor:
190        mod = self.adaLN_modulation(vec)
191        shift, scale = mod.chunk(2, dim=-1)
192        if shift.ndim == 2:
193            shift = shift[:, None, :]
194            scale = scale[:, None, :]
195        x = (1 + scale) * self.norm_final(x) + shift
196        x = self.linear(x)
197        return x
198
199
200class SingleStreamBlock(nn.Module):
201    def __init__(
202        self,
203        hidden_size: int,
204        num_heads: int,
205        mlp_ratio: float = 4.0,
206    ):
207        super().__init__()
208
209        self.hidden_dim = hidden_size
210        self.num_heads = num_heads
211        head_dim = hidden_size // num_heads
212        self.scale = head_dim**-0.5
213        self.mlp_hidden_dim = int(hidden_size * mlp_ratio)
214        self.mlp_mult_factor = 2
215
216        self.linear1 = nn.Linear(
217            hidden_size,
218            hidden_size * 3 + self.mlp_hidden_dim * self.mlp_mult_factor,
219            bias=False,
220        )
221
222        self.linear2 = nn.Linear(hidden_size + self.mlp_hidden_dim, hidden_size, bias=False)
223
224        self.norm = QKNorm(head_dim)
225
226        self.hidden_size = hidden_size
227        self.pre_norm = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
228
229        self.mlp_act = SiLUActivation()
230
231    def forward(
232        self,
233        x: Tensor,
234        pe: Tensor,
235        mod: tuple[Tensor, Tensor],
236    ) -> Tensor:
237        mod_shift, mod_scale, mod_gate = mod  # type: ignore
238        x_mod = (1 + mod_scale) * self.pre_norm(x) + mod_shift
239
240        qkv, mlp = torch.split(
241            self.linear1(x_mod),
242            [3 * self.hidden_size, self.mlp_hidden_dim * self.mlp_mult_factor],
243            dim=-1,
244        )
245
246        q, k, v = rearrange(qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
247        q, k = self.norm(q, k, v)
248
249        attn = attention(q, k, v, pe)
250
251        # compute activation in mlp stream, cat again and run second linear layer
252        output = self.linear2(torch.cat((attn, self.mlp_act(mlp)), 2))
253        return x + mod_gate * output
254
255
256class DoubleStreamBlock(nn.Module):
257    def __init__(
258        self,
259        hidden_size: int,
260        num_heads: int,
261        mlp_ratio: float,
262    ):
263        super().__init__()
264        mlp_hidden_dim = int(hidden_size * mlp_ratio)
265        self.num_heads = num_heads
266        assert hidden_size % num_heads == 0, f"{hidden_size=} must be divisible by {num_heads=}"
267
268        self.hidden_size = hidden_size
269        self.img_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
270        self.mlp_mult_factor = 2
271
272        self.img_attn = SelfAttention(
273            dim=hidden_size,
274            num_heads=num_heads,
275        )
276
277        self.img_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
278        self.img_mlp = nn.Sequential(
279            nn.Linear(hidden_size, mlp_hidden_dim * self.mlp_mult_factor, bias=False),
280            SiLUActivation(),
281            nn.Linear(mlp_hidden_dim, hidden_size, bias=False),
282        )
283
284        self.txt_norm1 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
285        self.txt_attn = SelfAttention(
286            dim=hidden_size,
287            num_heads=num_heads,
288        )
289
290        self.txt_norm2 = nn.LayerNorm(hidden_size, elementwise_affine=False, eps=1e-6)
291        self.txt_mlp = nn.Sequential(
292            nn.Linear(
293                hidden_size,
294                mlp_hidden_dim * self.mlp_mult_factor,
295                bias=False,
296            ),
297            SiLUActivation(),
298            nn.Linear(mlp_hidden_dim, hidden_size, bias=False),
299        )
300
301    def forward(
302        self,
303        img: Tensor,
304        txt: Tensor,
305        pe: Tensor,
306        pe_ctx: Tensor,
307        mod_img: tuple[Tensor, Tensor],
308        mod_txt: tuple[Tensor, Tensor],
309    ) -> tuple[Tensor, Tensor]:
310        img_mod1, img_mod2 = mod_img
311        txt_mod1, txt_mod2 = mod_txt
312
313        img_mod1_shift, img_mod1_scale, img_mod1_gate = img_mod1
314        img_mod2_shift, img_mod2_scale, img_mod2_gate = img_mod2
315        txt_mod1_shift, txt_mod1_scale, txt_mod1_gate = txt_mod1
316        txt_mod2_shift, txt_mod2_scale, txt_mod2_gate = txt_mod2
317
318        # prepare image for attention
319        img_modulated = self.img_norm1(img)
320        img_modulated = (1 + img_mod1_scale) * img_modulated + img_mod1_shift
321
322        img_qkv = self.img_attn.qkv(img_modulated)
323        img_q, img_k, img_v = rearrange(img_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
324        img_q, img_k = self.img_attn.norm(img_q, img_k, img_v)
325
326        # prepare txt for attention
327        txt_modulated = self.txt_norm1(txt)
328        txt_modulated = (1 + txt_mod1_scale) * txt_modulated + txt_mod1_shift
329
330        txt_qkv = self.txt_attn.qkv(txt_modulated)
331        txt_q, txt_k, txt_v = rearrange(txt_qkv, "B L (K H D) -> K B H L D", K=3, H=self.num_heads)
332        txt_q, txt_k = self.txt_attn.norm(txt_q, txt_k, txt_v)
333
334        q = torch.cat((txt_q, img_q), dim=2)
335        k = torch.cat((txt_k, img_k), dim=2)
336        v = torch.cat((txt_v, img_v), dim=2)
337
338        pe = torch.cat((pe_ctx, pe), dim=2)
339        attn = attention(q, k, v, pe)
340        txt_attn, img_attn = attn[:, : txt_q.shape[2]], attn[:, txt_q.shape[2] :]
341
342        # calculate the img blocks
343        img = img + img_mod1_gate * self.img_attn.proj(img_attn)
344        img = img + img_mod2_gate * self.img_mlp((1 + img_mod2_scale) * (self.img_norm2(img)) + img_mod2_shift)
345
346        # calculate the txt blocks
347        txt = txt + txt_mod1_gate * self.txt_attn.proj(txt_attn)
348        txt = txt + txt_mod2_gate * self.txt_mlp((1 + txt_mod2_scale) * (self.txt_norm2(txt)) + txt_mod2_shift)
349        return img, txt
350
351
352class MLPEmbedder(nn.Module):
353    def __init__(self, in_dim: int, hidden_dim: int, disable_bias: bool = False):
354        super().__init__()
355        self.in_layer = nn.Linear(in_dim, hidden_dim, bias=not disable_bias)
356        self.silu = nn.SiLU()
357        self.out_layer = nn.Linear(hidden_dim, hidden_dim, bias=not disable_bias)
358
359    def forward(self, x: Tensor) -> Tensor:
360        return self.out_layer(self.silu(self.in_layer(x)))
361
362
363class EmbedND(nn.Module):
364    def __init__(self, dim: int, theta: int, axes_dim: list[int]):
365        super().__init__()
366        self.dim = dim
367        self.theta = theta
368        self.axes_dim = axes_dim
369
370    def forward(self, ids: Tensor) -> Tensor:
371        emb = torch.cat(
372            [rope(ids[..., i], self.axes_dim[i], self.theta) for i in range(len(self.axes_dim))],
373            dim=-3,
374        )
375
376        return emb.unsqueeze(1)
377
378
379def timestep_embedding(t: Tensor, dim, max_period=10000, time_factor: float = 1000.0):
380    """
381    Create sinusoidal timestep embeddings.
382    :param t: a 1-D Tensor of N indices, one per batch element.
383                      These may be fractional.
384    :param dim: the dimension of the output.
385    :param max_period: controls the minimum frequency of the embeddings.
386    :return: an (N, D) Tensor of positional embeddings.
387    """
388    t = time_factor * t
389    half = dim // 2
390    freqs = torch.exp(
391        -math.log(max_period) * torch.arange(start=0, end=half, device=t.device, dtype=torch.float32) / half  # float32 originally
392    )
393
394    args = t[:, None].float() * freqs[None]
395    embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
396    if dim % 2:
397        embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
398    if torch.is_floating_point(t):
399        embedding = embedding.to(t)
400    return embedding
401
402
403class RMSNorm(torch.nn.Module):
404    def __init__(self, dim: int):
405        super().__init__()
406        self.scale = nn.Parameter(torch.ones(dim))
407
408    def forward(self, x: Tensor):
409        x_dtype = x.dtype
410        x = x.float()
411        rrms = torch.rsqrt(torch.mean(x**2, dim=-1, keepdim=True) + 1e-6)
412        return (x * rrms).to(dtype=x_dtype) * self.scale
413
414
415class QKNorm(torch.nn.Module):
416    def __init__(self, dim: int):
417        super().__init__()
418        self.query_norm = RMSNorm(dim)
419        self.key_norm = RMSNorm(dim)
420
421    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> tuple[Tensor, Tensor]:
422        q = self.query_norm(q)
423        k = self.key_norm(k)
424        return q.to(v), k.to(v)
425
426
427def attention(q: Tensor, k: Tensor, v: Tensor, pe: Tensor) -> Tensor:
428    q, k = apply_rope(q, k, pe)
429
430    x = torch.nn.functional.scaled_dot_product_attention(q, k, v)
431    x = rearrange(x, "B H L D -> B L (H D)")
432
433    return x
434
435
436def rope(pos: Tensor, dim: int, theta: int) -> Tensor:
437    assert dim % 2 == 0
438    scale = torch.arange(0, dim, 2, dtype=pos.dtype, device=pos.device) / dim
439    omega = 1.0 / (theta**scale)
440    out = torch.einsum("...n,d->...nd", pos, omega)
441    out = torch.stack([torch.cos(out), -torch.sin(out), torch.sin(out), torch.cos(out)], dim=-1)
442    out = rearrange(out, "b n d (i j) -> b n d i j", i=2, j=2)
443    return out.float()
444
445
446def apply_rope(xq: Tensor, xk: Tensor, freqs_cis: Tensor) -> tuple[Tensor, Tensor]:
447    xq_ = xq.float().reshape(*xq.shape[:-1], -1, 1, 2)
448    xk_ = xk.float().reshape(*xk.shape[:-1], -1, 1, 2)
449    xq_out = freqs_cis[..., 0] * xq_[..., 0] + freqs_cis[..., 1] * xq_[..., 1]
450    xk_out = freqs_cis[..., 0] * xk_[..., 0] + freqs_cis[..., 1] * xk_[..., 1]
451    return xq_out.reshape(*xq.shape).type_as(xq), xk_out.reshape(*xk.shape).type_as(xk)