attention.py 27.2 KB
Newer Older
comfyanonymous's avatar
comfyanonymous committed
1
2
3
4
5
6
7
from inspect import isfunction
import math
import torch
import torch.nn.functional as F
from torch import nn, einsum
from einops import rearrange, repeat
from typing import Optional, Any
comfyanonymous's avatar
comfyanonymous committed
8
from functools import partial
comfyanonymous's avatar
comfyanonymous committed
9

comfyanonymous's avatar
comfyanonymous committed
10
11

from .diffusionmodules.util import checkpoint, AlphaBlender, timestep_embedding
comfyanonymous's avatar
comfyanonymous committed
12
13
from .sub_quadratic_attention import efficient_dot_product_attention

14
from comfy import model_management
15

16
if model_management.xformers_enabled():
comfyanonymous's avatar
comfyanonymous committed
17
18
19
    import xformers
    import xformers.ops

comfyanonymous's avatar
comfyanonymous committed
20
from comfy.cli_args import args
comfyanonymous's avatar
comfyanonymous committed
21
22
import comfy.ops

comfyanonymous's avatar
comfyanonymous committed
23
# CrossAttn precision handling
comfyanonymous's avatar
comfyanonymous committed
24
25
26
27
28
if args.dont_upcast_attention:
    print("disabling upcasting of attention")
    _ATTN_PRECISION = "fp16"
else:
    _ATTN_PRECISION = "fp32"
comfyanonymous's avatar
comfyanonymous committed
29

30

comfyanonymous's avatar
comfyanonymous committed
31
32
33
34
35
36
37
38
39
40
41
def exists(val):
    return val is not None


def uniq(arr):
    return{el: True for el in arr}.keys()


def default(val, d):
    if exists(val):
        return val
42
    return d
comfyanonymous's avatar
comfyanonymous committed
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57


def max_neg_value(t):
    return -torch.finfo(t.dtype).max


def init_(tensor):
    dim = tensor.shape[-1]
    std = 1 / math.sqrt(dim)
    tensor.uniform_(-std, std)
    return tensor


# feedforward
class GEGLU(nn.Module):
comfyanonymous's avatar
comfyanonymous committed
58
    def __init__(self, dim_in, dim_out, dtype=None, device=None, operations=comfy.ops):
comfyanonymous's avatar
comfyanonymous committed
59
        super().__init__()
comfyanonymous's avatar
comfyanonymous committed
60
        self.proj = operations.Linear(dim_in, dim_out * 2, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
61
62
63
64
65
66
67

    def forward(self, x):
        x, gate = self.proj(x).chunk(2, dim=-1)
        return x * F.gelu(gate)


class FeedForward(nn.Module):
comfyanonymous's avatar
comfyanonymous committed
68
    def __init__(self, dim, dim_out=None, mult=4, glu=False, dropout=0., dtype=None, device=None, operations=comfy.ops):
comfyanonymous's avatar
comfyanonymous committed
69
70
71
72
        super().__init__()
        inner_dim = int(dim * mult)
        dim_out = default(dim_out, dim)
        project_in = nn.Sequential(
comfyanonymous's avatar
comfyanonymous committed
73
            operations.Linear(dim, inner_dim, dtype=dtype, device=device),
comfyanonymous's avatar
comfyanonymous committed
74
            nn.GELU()
comfyanonymous's avatar
comfyanonymous committed
75
        ) if not glu else GEGLU(dim, inner_dim, dtype=dtype, device=device, operations=operations)
comfyanonymous's avatar
comfyanonymous committed
76
77
78
79

        self.net = nn.Sequential(
            project_in,
            nn.Dropout(dropout),
comfyanonymous's avatar
comfyanonymous committed
80
            operations.Linear(inner_dim, dim_out, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
81
82
83
84
85
        )

    def forward(self, x):
        return self.net(x)

86
87
def Normalize(in_channels, dtype=None, device=None):
    return torch.nn.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
88

89
def attention_basic(q, k, v, heads, mask=None):
90
91
92
93
    b, _, dim_head = q.shape
    dim_head //= heads
    scale = dim_head ** -0.5

94
    h = heads
95
96
97
98
99
100
101
102
    q, k, v = map(
        lambda t: t.unsqueeze(3)
        .reshape(b, -1, heads, dim_head)
        .permute(0, 2, 1, 3)
        .reshape(b * heads, -1, dim_head)
        .contiguous(),
        (q, k, v),
    )
103
104
105
106
107
108
109
110

    # force cast to fp32 to avoid overflowing
    if _ATTN_PRECISION =="fp32":
        with torch.autocast(enabled=False, device_type = 'cuda'):
            q, k = q.float(), k.float()
            sim = einsum('b i d, b j d -> b i j', q, k) * scale
    else:
        sim = einsum('b i d, b j d -> b i j', q, k) * scale
comfyanonymous's avatar
comfyanonymous committed
111

112
    del q, k
comfyanonymous's avatar
comfyanonymous committed
113

114
    if exists(mask):
115
116
117
118
119
120
121
        if mask.dtype == torch.bool:
            mask = rearrange(mask, 'b ... -> b (...)') #TODO: check if this bool part matches pytorch attention
            max_neg_value = -torch.finfo(sim.dtype).max
            mask = repeat(mask, 'b j -> (b h) () j', h=h)
            sim.masked_fill_(~mask, max_neg_value)
        else:
            sim += mask
comfyanonymous's avatar
comfyanonymous committed
122

123
124
    # attention, what we cannot get enough of
    sim = sim.softmax(dim=-1)
comfyanonymous's avatar
comfyanonymous committed
125

126
    out = einsum('b i j, b j d -> b i d', sim.to(v.dtype), v)
127
128
129
130
131
132
    out = (
        out.unsqueeze(0)
        .reshape(b, heads, -1, dim_head)
        .permute(0, 2, 1, 3)
        .reshape(b, -1, heads * dim_head)
    )
133
    return out
comfyanonymous's avatar
comfyanonymous committed
134
135


136
def attention_sub_quad(query, key, value, heads, mask=None):
137
138
139
140
141
142
143
144
    b, _, dim_head = query.shape
    dim_head //= heads

    scale = dim_head ** -0.5
    query = query.unsqueeze(3).reshape(b, -1, heads, dim_head).permute(0, 2, 1, 3).reshape(b * heads, -1, dim_head)
    value = value.unsqueeze(3).reshape(b, -1, heads, dim_head).permute(0, 2, 1, 3).reshape(b * heads, -1, dim_head)

    key = key.unsqueeze(3).reshape(b, -1, heads, dim_head).permute(0, 2, 3, 1).reshape(b * heads, dim_head, -1)
comfyanonymous's avatar
comfyanonymous committed
145

146
147
148
149
150
151
152
    dtype = query.dtype
    upcast_attention = _ATTN_PRECISION =="fp32" and query.dtype != torch.float32
    if upcast_attention:
        bytes_per_token = torch.finfo(torch.float32).bits//8
    else:
        bytes_per_token = torch.finfo(query.dtype).bits//8
    batch_x_heads, q_tokens, _ = query.shape
153
    _, _, k_tokens = key.shape
154
    qk_matmul_size_bytes = batch_x_heads * bytes_per_token * q_tokens * k_tokens
comfyanonymous's avatar
comfyanonymous committed
155

156
    mem_free_total, mem_free_torch = model_management.get_free_memory(query.device, True)
comfyanonymous's avatar
comfyanonymous committed
157

158
    kv_chunk_size_min = None
159
160
161
162
163
164
165
166
167
168
169
170
    kv_chunk_size = None
    query_chunk_size = None

    for x in [4096, 2048, 1024, 512, 256]:
        count = mem_free_total / (batch_x_heads * bytes_per_token * x * 4.0)
        if count >= k_tokens:
            kv_chunk_size = k_tokens
            query_chunk_size = x
            break

    if query_chunk_size is None:
        query_chunk_size = 512
171
172
173

    hidden_states = efficient_dot_product_attention(
        query,
174
        key,
175
176
177
178
179
180
181
182
183
184
185
186
187
188
        value,
        query_chunk_size=query_chunk_size,
        kv_chunk_size=kv_chunk_size,
        kv_chunk_size_min=kv_chunk_size_min,
        use_checkpoint=False,
        upcast_attention=upcast_attention,
    )

    hidden_states = hidden_states.to(dtype)

    hidden_states = hidden_states.unflatten(0, (-1, heads)).transpose(1,2).flatten(start_dim=2)
    return hidden_states

def attention_split(q, k, v, heads, mask=None):
189
190
191
192
    b, _, dim_head = q.shape
    dim_head //= heads
    scale = dim_head ** -0.5

193
    h = heads
194
195
196
197
198
199
200
201
    q, k, v = map(
        lambda t: t.unsqueeze(3)
        .reshape(b, -1, heads, dim_head)
        .permute(0, 2, 1, 3)
        .reshape(b * heads, -1, dim_head)
        .contiguous(),
        (q, k, v),
    )
202
203
204
205
206

    r1 = torch.zeros(q.shape[0], q.shape[1], v.shape[2], device=q.device, dtype=q.dtype)

    mem_free_total = model_management.get_free_memory(q.device)

207
208
209
210
211
    if _ATTN_PRECISION =="fp32":
        element_size = 4
    else:
        element_size = q.element_size()

212
    gb = 1024 ** 3
213
    tensor_size = q.shape[0] * q.shape[1] * k.shape[1] * element_size
214
    modifier = 3
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
    mem_required = tensor_size * modifier
    steps = 1


    if mem_required > mem_free_total:
        steps = 2**(math.ceil(math.log(mem_required / mem_free_total, 2)))
        # print(f"Expected tensor size:{tensor_size/gb:0.1f}GB, cuda free:{mem_free_cuda/gb:0.1f}GB "
        #      f"torch free:{mem_free_torch/gb:0.1f} total:{mem_free_total/gb:0.1f} steps:{steps}")

    if steps > 64:
        max_res = math.floor(math.sqrt(math.sqrt(mem_free_total / 2.5)) / 8) * 64
        raise RuntimeError(f'Not enough memory, use lower resolution (max approx. {max_res}x{max_res}). '
                            f'Need: {mem_required/64/gb:0.1f}GB free, Have:{mem_free_total/gb:0.1f}GB free')

    # print("steps", steps, mem_required, mem_free_total, modifier, q.element_size(), tensor_size)
    first_op_done = False
    cleared_cache = False
    while True:
        try:
            slice_size = q.shape[1] // steps if (q.shape[1] % steps) == 0 else q.shape[1]
            for i in range(0, q.shape[1], slice_size):
                end = i + slice_size
                if _ATTN_PRECISION =="fp32":
                    with torch.autocast(enabled=False, device_type = 'cuda'):
                        s1 = einsum('b i d, b j d -> b i j', q[:, i:end].float(), k.float()) * scale
comfyanonymous's avatar
comfyanonymous committed
240
                else:
241
242
243
244
                    s1 = einsum('b i d, b j d -> b i j', q[:, i:end], k) * scale

                s2 = s1.softmax(dim=-1).to(v.dtype)
                del s1
245
                first_op_done = True
246
247
248
249
250
251
252
253
254
255
256
257
258

                r1[:, i:end] = einsum('b i j, b j d -> b i d', s2, v)
                del s2
            break
        except model_management.OOM_EXCEPTION as e:
            if first_op_done == False:
                model_management.soft_empty_cache(True)
                if cleared_cache == False:
                    cleared_cache = True
                    print("out of memory error, emptying cache and trying again")
                    continue
                steps *= 2
                if steps > 64:
comfyanonymous's avatar
comfyanonymous committed
259
                    raise e
260
261
262
263
264
265
                print("out of memory error, increasing steps and trying again", steps)
            else:
                raise e

    del q, k, v

266
267
268
269
270
271
272
    r1 = (
        r1.unsqueeze(0)
        .reshape(b, heads, -1, dim_head)
        .permute(0, 2, 1, 3)
        .reshape(b, -1, heads * dim_head)
    )
    return r1
273

274
275
276
277
278
279
280
281
BROKEN_XFORMERS = False
try:
    x_vers = xformers.__version__
    #I think 0.0.23 is also broken (q with bs bigger than 65535 gives CUDA error)
    BROKEN_XFORMERS = x_vers.startswith("0.0.21") or x_vers.startswith("0.0.22") or x_vers.startswith("0.0.23")
except:
    pass

282
def attention_xformers(q, k, v, heads, mask=None):
283
284
    b, _, dim_head = q.shape
    dim_head //= heads
285
286
287
    if BROKEN_XFORMERS:
        if b * heads > 65535:
            return attention_pytorch(q, k, v, heads, mask)
288

289
290
    q, k, v = map(
        lambda t: t.unsqueeze(3)
291
        .reshape(b, -1, heads, dim_head)
292
        .permute(0, 2, 1, 3)
293
        .reshape(b * heads, -1, dim_head)
294
295
296
297
298
299
300
301
302
303
304
        .contiguous(),
        (q, k, v),
    )

    # actually compute the attention, what we cannot get enough of
    out = xformers.ops.memory_efficient_attention(q, k, v, attn_bias=None)

    if exists(mask):
        raise NotImplementedError
    out = (
        out.unsqueeze(0)
305
        .reshape(b, heads, -1, dim_head)
306
        .permute(0, 2, 1, 3)
307
        .reshape(b, -1, heads * dim_head)
308
309
310
311
312
313
314
315
316
317
318
    )
    return out

def attention_pytorch(q, k, v, heads, mask=None):
    b, _, dim_head = q.shape
    dim_head //= heads
    q, k, v = map(
        lambda t: t.view(b, -1, heads, dim_head).transpose(1, 2),
        (q, k, v),
    )

319
    out = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=mask, dropout_p=0.0, is_causal=False)
320
321
322
323
324
    out = (
        out.transpose(1, 2).reshape(b, -1, heads * dim_head)
    )
    return out

325

326
optimized_attention = attention_basic
327
optimized_attention_masked = attention_basic
comfyanonymous's avatar
comfyanonymous committed
328

329
330
331
332
333
334
335
336
337
338
339
340
341
if model_management.xformers_enabled():
    print("Using xformers cross attention")
    optimized_attention = attention_xformers
elif model_management.pytorch_attention_enabled():
    print("Using pytorch cross attention")
    optimized_attention = attention_pytorch
else:
    if args.use_split_cross_attention:
        print("Using split optimization for cross attention")
        optimized_attention = attention_split
    else:
        print("Using sub quadratic optimization for cross attention, if you have memory or speed issues try using: --use-split-cross-attention")
        optimized_attention = attention_sub_quad
comfyanonymous's avatar
comfyanonymous committed
342

343
344
345
if model_management.pytorch_attention_enabled():
    optimized_attention_masked = attention_pytorch

346
347
348
349
350
351
352
353
354
355
356
357
def optimized_attention_for_device(device, mask=False):
    if device == torch.device("cpu"): #TODO
        if model_management.pytorch_attention_enabled():
            return attention_pytorch
        else:
            return attention_basic
    if mask:
        return optimized_attention_masked

    return optimized_attention


358
class CrossAttention(nn.Module):
comfyanonymous's avatar
comfyanonymous committed
359
    def __init__(self, query_dim, context_dim=None, heads=8, dim_head=64, dropout=0., dtype=None, device=None, operations=comfy.ops):
360
361
362
363
364
365
366
        super().__init__()
        inner_dim = dim_head * heads
        context_dim = default(context_dim, query_dim)

        self.heads = heads
        self.dim_head = dim_head

comfyanonymous's avatar
comfyanonymous committed
367
368
369
        self.to_q = operations.Linear(query_dim, inner_dim, bias=False, dtype=dtype, device=device)
        self.to_k = operations.Linear(context_dim, inner_dim, bias=False, dtype=dtype, device=device)
        self.to_v = operations.Linear(context_dim, inner_dim, bias=False, dtype=dtype, device=device)
370

comfyanonymous's avatar
comfyanonymous committed
371
        self.to_out = nn.Sequential(operations.Linear(inner_dim, query_dim, dtype=dtype, device=device), nn.Dropout(dropout))
372

373
    def forward(self, x, context=None, value=None, mask=None):
374
375
376
        q = self.to_q(x)
        context = default(context, x)
        k = self.to_k(context)
377
378
379
380
381
        if value is not None:
            v = self.to_v(value)
            del value
        else:
            v = self.to_v(context)
382

383
384
385
386
        if mask is None:
            out = optimized_attention(q, k, v, self.heads)
        else:
            out = optimized_attention_masked(q, k, v, self.heads, mask)
387
388
        return self.to_out(out)

389

comfyanonymous's avatar
comfyanonymous committed
390
class BasicTransformerBlock(nn.Module):
comfyanonymous's avatar
comfyanonymous committed
391
392
    def __init__(self, dim, n_heads, d_head, dropout=0., context_dim=None, gated_ff=True, checkpoint=True, ff_in=False, inner_dim=None,
                 disable_self_attn=False, disable_temporal_crossattention=False, switch_temporal_ca_to_sa=False, dtype=None, device=None, operations=comfy.ops):
comfyanonymous's avatar
comfyanonymous committed
393
        super().__init__()
comfyanonymous's avatar
comfyanonymous committed
394
395
396
397
398
399
400
401

        self.ff_in = ff_in or inner_dim is not None
        if inner_dim is None:
            inner_dim = dim

        self.is_res = inner_dim == dim

        if self.ff_in:
comfyanonymous's avatar
comfyanonymous committed
402
            self.norm_in = operations.LayerNorm(dim, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
403
404
            self.ff_in = FeedForward(dim, dim_out=inner_dim, dropout=dropout, glu=gated_ff, dtype=dtype, device=device, operations=operations)

comfyanonymous's avatar
comfyanonymous committed
405
        self.disable_self_attn = disable_self_attn
comfyanonymous's avatar
comfyanonymous committed
406
        self.attn1 = CrossAttention(query_dim=inner_dim, heads=n_heads, dim_head=d_head, dropout=dropout,
comfyanonymous's avatar
comfyanonymous committed
407
                              context_dim=context_dim if self.disable_self_attn else None, dtype=dtype, device=device, operations=operations)  # is a self-attention if not self.disable_self_attn
comfyanonymous's avatar
comfyanonymous committed
408
409
410
411
412
413
414
415
416
417
418
419
420
421
        self.ff = FeedForward(inner_dim, dim_out=dim, dropout=dropout, glu=gated_ff, dtype=dtype, device=device, operations=operations)

        if disable_temporal_crossattention:
            if switch_temporal_ca_to_sa:
                raise ValueError
            else:
                self.attn2 = None
        else:
            context_dim_attn2 = None
            if not switch_temporal_ca_to_sa:
                context_dim_attn2 = context_dim

            self.attn2 = CrossAttention(query_dim=inner_dim, context_dim=context_dim_attn2,
                                heads=n_heads, dim_head=d_head, dropout=dropout, dtype=dtype, device=device, operations=operations)  # is self-attn if context is none
422
            self.norm2 = operations.LayerNorm(inner_dim, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
423

424
425
        self.norm1 = operations.LayerNorm(inner_dim, dtype=dtype, device=device)
        self.norm3 = operations.LayerNorm(inner_dim, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
426
        self.checkpoint = checkpoint
427
428
        self.n_heads = n_heads
        self.d_head = d_head
comfyanonymous's avatar
comfyanonymous committed
429
        self.switch_temporal_ca_to_sa = switch_temporal_ca_to_sa
comfyanonymous's avatar
comfyanonymous committed
430

431
432
    def forward(self, x, context=None, transformer_options={}):
        return checkpoint(self._forward, (x, context, transformer_options), self.parameters(), self.checkpoint)
comfyanonymous's avatar
comfyanonymous committed
433

434
    def _forward(self, x, context=None, transformer_options={}):
435
        extra_options = {}
436
437
        block = transformer_options.get("block", None)
        block_index = transformer_options.get("block_index", 0)
438
439
440
441
442
443
444
445
446
447
        transformer_patches = {}
        transformer_patches_replace = {}

        for k in transformer_options:
            if k == "patches":
                transformer_patches = transformer_options[k]
            elif k == "patches_replace":
                transformer_patches_replace = transformer_options[k]
            else:
                extra_options[k] = transformer_options[k]
448

449
450
451
        extra_options["n_heads"] = self.n_heads
        extra_options["dim_head"] = self.d_head

comfyanonymous's avatar
comfyanonymous committed
452
453
454
455
456
457
        if self.ff_in:
            x_skip = x
            x = self.ff_in(self.norm_in(x))
            if self.is_res:
                x += x_skip

458
        n = self.norm1(x)
459
460
461
462
463
464
465
466
467
468
469
470
        if self.disable_self_attn:
            context_attn1 = context
        else:
            context_attn1 = None
        value_attn1 = None

        if "attn1_patch" in transformer_patches:
            patch = transformer_patches["attn1_patch"]
            if context_attn1 is None:
                context_attn1 = n
            value_attn1 = context_attn1
            for p in patch:
471
                n, context_attn1, value_attn1 = p(n, context_attn1, value_attn1, extra_options)
472

comfyanonymous's avatar
comfyanonymous committed
473
474
475
476
        if block is not None:
            transformer_block = (block[0], block[1], block_index)
        else:
            transformer_block = None
477
478
479
480
481
482
483
484
485
486
487
488
489
490
        attn1_replace_patch = transformer_patches_replace.get("attn1", {})
        block_attn1 = transformer_block
        if block_attn1 not in attn1_replace_patch:
            block_attn1 = block

        if block_attn1 in attn1_replace_patch:
            if context_attn1 is None:
                context_attn1 = n
                value_attn1 = n
            n = self.attn1.to_q(n)
            context_attn1 = self.attn1.to_k(context_attn1)
            value_attn1 = self.attn1.to_v(value_attn1)
            n = attn1_replace_patch[block_attn1](n, context_attn1, value_attn1, extra_options)
            n = self.attn1.to_out(n)
491
        else:
492
            n = self.attn1(n, context=context_attn1, value=value_attn1)
493

494
495
496
497
498
        if "attn1_output_patch" in transformer_patches:
            patch = transformer_patches["attn1_output_patch"]
            for p in patch:
                n = p(n, extra_options)

499
        x += n
500
501
502
        if "middle_patch" in transformer_patches:
            patch = transformer_patches["middle_patch"]
            for p in patch:
503
                x = p(x, extra_options)
504

comfyanonymous's avatar
comfyanonymous committed
505
506
507
508
509
510
511
512
513
        if self.attn2 is not None:
            n = self.norm2(x)
            if self.switch_temporal_ca_to_sa:
                context_attn2 = n
            else:
                context_attn2 = context
            value_attn2 = None
            if "attn2_patch" in transformer_patches:
                patch = transformer_patches["attn2_patch"]
514
                value_attn2 = context_attn2
comfyanonymous's avatar
comfyanonymous committed
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
                for p in patch:
                    n, context_attn2, value_attn2 = p(n, context_attn2, value_attn2, extra_options)

            attn2_replace_patch = transformer_patches_replace.get("attn2", {})
            block_attn2 = transformer_block
            if block_attn2 not in attn2_replace_patch:
                block_attn2 = block

            if block_attn2 in attn2_replace_patch:
                if value_attn2 is None:
                    value_attn2 = context_attn2
                n = self.attn2.to_q(n)
                context_attn2 = self.attn2.to_k(context_attn2)
                value_attn2 = self.attn2.to_v(value_attn2)
                n = attn2_replace_patch[block_attn2](n, context_attn2, value_attn2, extra_options)
                n = self.attn2.to_out(n)
            else:
                n = self.attn2(n, context=context_attn2, value=value_attn2)
533

534
535
536
537
538
        if "attn2_output_patch" in transformer_patches:
            patch = transformer_patches["attn2_output_patch"]
            for p in patch:
                n = p(n, extra_options)

539
        x += n
comfyanonymous's avatar
comfyanonymous committed
540
541
542
543
544
545
        if self.is_res:
            x_skip = x
        x = self.ff(self.norm3(x))
        if self.is_res:
            x += x_skip

comfyanonymous's avatar
comfyanonymous committed
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
        return x


class SpatialTransformer(nn.Module):
    """
    Transformer block for image-like data.
    First, project the input (aka embedding)
    and reshape to b, t, d.
    Then apply standard transformer action.
    Finally, reshape to image
    NEW: use_linear for more efficiency instead of the 1x1 convs
    """
    def __init__(self, in_channels, n_heads, d_head,
                 depth=1, dropout=0., context_dim=None,
                 disable_self_attn=False, use_linear=False,
comfyanonymous's avatar
comfyanonymous committed
561
                 use_checkpoint=True, dtype=None, device=None, operations=comfy.ops):
comfyanonymous's avatar
comfyanonymous committed
562
563
        super().__init__()
        if exists(context_dim) and not isinstance(context_dim, list):
564
            context_dim = [context_dim] * depth
comfyanonymous's avatar
comfyanonymous committed
565
566
        self.in_channels = in_channels
        inner_dim = n_heads * d_head
567
        self.norm = operations.GroupNorm(num_groups=32, num_channels=in_channels, eps=1e-6, affine=True, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
568
        if not use_linear:
comfyanonymous's avatar
comfyanonymous committed
569
            self.proj_in = operations.Conv2d(in_channels,
comfyanonymous's avatar
comfyanonymous committed
570
571
572
                                     inner_dim,
                                     kernel_size=1,
                                     stride=1,
573
                                     padding=0, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
574
        else:
comfyanonymous's avatar
comfyanonymous committed
575
            self.proj_in = operations.Linear(in_channels, inner_dim, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
576
577
578

        self.transformer_blocks = nn.ModuleList(
            [BasicTransformerBlock(inner_dim, n_heads, d_head, dropout=dropout, context_dim=context_dim[d],
comfyanonymous's avatar
comfyanonymous committed
579
                                   disable_self_attn=disable_self_attn, checkpoint=use_checkpoint, dtype=dtype, device=device, operations=operations)
comfyanonymous's avatar
comfyanonymous committed
580
581
582
                for d in range(depth)]
        )
        if not use_linear:
comfyanonymous's avatar
comfyanonymous committed
583
            self.proj_out = operations.Conv2d(inner_dim,in_channels,
comfyanonymous's avatar
comfyanonymous committed
584
585
                                                  kernel_size=1,
                                                  stride=1,
586
                                                  padding=0, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
587
        else:
comfyanonymous's avatar
comfyanonymous committed
588
            self.proj_out = operations.Linear(in_channels, inner_dim, dtype=dtype, device=device)
comfyanonymous's avatar
comfyanonymous committed
589
590
        self.use_linear = use_linear

591
    def forward(self, x, context=None, transformer_options={}):
comfyanonymous's avatar
comfyanonymous committed
592
593
        # note: if no context is given, cross-attention defaults to self-attention
        if not isinstance(context, list):
594
            context = [context] * len(self.transformer_blocks)
comfyanonymous's avatar
comfyanonymous committed
595
596
597
598
599
600
601
602
603
        b, c, h, w = x.shape
        x_in = x
        x = self.norm(x)
        if not self.use_linear:
            x = self.proj_in(x)
        x = rearrange(x, 'b c h w -> b (h w) c').contiguous()
        if self.use_linear:
            x = self.proj_in(x)
        for i, block in enumerate(self.transformer_blocks):
604
            transformer_options["block_index"] = i
605
            x = block(x, context=context[i], transformer_options=transformer_options)
comfyanonymous's avatar
comfyanonymous committed
606
607
608
609
610
611
612
        if self.use_linear:
            x = self.proj_out(x)
        x = rearrange(x, 'b (h w) c -> b c h w', h=h, w=w).contiguous()
        if not self.use_linear:
            x = self.proj_out(x)
        return x + x_in

comfyanonymous's avatar
comfyanonymous committed
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773

class SpatialVideoTransformer(SpatialTransformer):
    def __init__(
        self,
        in_channels,
        n_heads,
        d_head,
        depth=1,
        dropout=0.0,
        use_linear=False,
        context_dim=None,
        use_spatial_context=False,
        timesteps=None,
        merge_strategy: str = "fixed",
        merge_factor: float = 0.5,
        time_context_dim=None,
        ff_in=False,
        checkpoint=False,
        time_depth=1,
        disable_self_attn=False,
        disable_temporal_crossattention=False,
        max_time_embed_period: int = 10000,
        dtype=None, device=None, operations=comfy.ops
    ):
        super().__init__(
            in_channels,
            n_heads,
            d_head,
            depth=depth,
            dropout=dropout,
            use_checkpoint=checkpoint,
            context_dim=context_dim,
            use_linear=use_linear,
            disable_self_attn=disable_self_attn,
            dtype=dtype, device=device, operations=operations
        )
        self.time_depth = time_depth
        self.depth = depth
        self.max_time_embed_period = max_time_embed_period

        time_mix_d_head = d_head
        n_time_mix_heads = n_heads

        time_mix_inner_dim = int(time_mix_d_head * n_time_mix_heads)

        inner_dim = n_heads * d_head
        if use_spatial_context:
            time_context_dim = context_dim

        self.time_stack = nn.ModuleList(
            [
                BasicTransformerBlock(
                    inner_dim,
                    n_time_mix_heads,
                    time_mix_d_head,
                    dropout=dropout,
                    context_dim=time_context_dim,
                    # timesteps=timesteps,
                    checkpoint=checkpoint,
                    ff_in=ff_in,
                    inner_dim=time_mix_inner_dim,
                    disable_self_attn=disable_self_attn,
                    disable_temporal_crossattention=disable_temporal_crossattention,
                    dtype=dtype, device=device, operations=operations
                )
                for _ in range(self.depth)
            ]
        )

        assert len(self.time_stack) == len(self.transformer_blocks)

        self.use_spatial_context = use_spatial_context
        self.in_channels = in_channels

        time_embed_dim = self.in_channels * 4
        self.time_pos_embed = nn.Sequential(
            operations.Linear(self.in_channels, time_embed_dim, dtype=dtype, device=device),
            nn.SiLU(),
            operations.Linear(time_embed_dim, self.in_channels, dtype=dtype, device=device),
        )

        self.time_mixer = AlphaBlender(
            alpha=merge_factor, merge_strategy=merge_strategy
        )

    def forward(
        self,
        x: torch.Tensor,
        context: Optional[torch.Tensor] = None,
        time_context: Optional[torch.Tensor] = None,
        timesteps: Optional[int] = None,
        image_only_indicator: Optional[torch.Tensor] = None,
        transformer_options={}
    ) -> torch.Tensor:
        _, _, h, w = x.shape
        x_in = x
        spatial_context = None
        if exists(context):
            spatial_context = context

        if self.use_spatial_context:
            assert (
                context.ndim == 3
            ), f"n dims of spatial context should be 3 but are {context.ndim}"

            if time_context is None:
                time_context = context
            time_context_first_timestep = time_context[::timesteps]
            time_context = repeat(
                time_context_first_timestep, "b ... -> (b n) ...", n=h * w
            )
        elif time_context is not None and not self.use_spatial_context:
            time_context = repeat(time_context, "b ... -> (b n) ...", n=h * w)
            if time_context.ndim == 2:
                time_context = rearrange(time_context, "b c -> b 1 c")

        x = self.norm(x)
        if not self.use_linear:
            x = self.proj_in(x)
        x = rearrange(x, "b c h w -> b (h w) c")
        if self.use_linear:
            x = self.proj_in(x)

        num_frames = torch.arange(timesteps, device=x.device)
        num_frames = repeat(num_frames, "t -> b t", b=x.shape[0] // timesteps)
        num_frames = rearrange(num_frames, "b t -> (b t)")
        t_emb = timestep_embedding(num_frames, self.in_channels, repeat_only=False, max_period=self.max_time_embed_period).to(x.dtype)
        emb = self.time_pos_embed(t_emb)
        emb = emb[:, None, :]

        for it_, (block, mix_block) in enumerate(
            zip(self.transformer_blocks, self.time_stack)
        ):
            transformer_options["block_index"] = it_
            x = block(
                x,
                context=spatial_context,
                transformer_options=transformer_options,
            )

            x_mix = x
            x_mix = x_mix + emb

            B, S, C = x_mix.shape
            x_mix = rearrange(x_mix, "(b t) s c -> (b s) t c", t=timesteps)
            x_mix = mix_block(x_mix, context=time_context) #TODO: transformer_options
            x_mix = rearrange(
                x_mix, "(b s) t c -> (b t) s c", s=S, b=B // timesteps, c=C, t=timesteps
            )

            x = self.time_mixer(x_spatial=x, x_temporal=x_mix, image_only_indicator=image_only_indicator)

        if self.use_linear:
            x = self.proj_out(x)
        x = rearrange(x, "b (h w) c -> b c h w", h=h, w=w)
        if not self.use_linear:
            x = self.proj_out(x)
        out = x + x_in
        return out