File size: 19,206 Bytes
3d1c0e1
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) 2025 FoundationVision
# SPDX-License-Identifier: MIT

"""
Definitions of blocks of VAR transformer model.
"""

import math
import os
from functools import partial
from typing import Optional, Tuple, Union

import torch
import torch.nn as nn
import torch.nn.functional as F
import numpy as np

from infinity.models.rope import apply_rotary_emb
from infinity.utils.sequence_parallel import sp_all_to_all, SequenceParallelManager as sp_manager

# Import flash_attn's fused ops
try:
    from flash_attn.ops.rms_norm import rms_norm as rms_norm_impl
    from flash_attn.ops.fused_dense import fused_mlp_func
    flash_fused_op_installed = True
except ImportError:
    fused_mlp_func = None
    flash_fused_op_installed = False
    
    def rms_norm_impl(x, weight, epsilon):
        return (x * torch.rsqrt(x.pow(2).mean(-1, keepdim=True).add_(epsilon))) * weight


class FastRMSNorm(nn.Module):
    def __init__(self, C, eps=1e-6, elementwise_affine=True):
        super().__init__()
        self.C = C
        self.eps = eps
        self.elementwise_affine = elementwise_affine
        if self.elementwise_affine:
            self.weight = nn.Parameter(torch.ones(C))
        else:
            self.register_buffer('weight', torch.ones(C))
    
    def forward(self, x):
        src_type = x.dtype
        return rms_norm_impl(x.float(), self.weight, epsilon=self.eps).to(src_type)
    
    def extra_repr(self) -> str:
        return f'C={self.C}, eps={self.eps:g}, elementwise_affine={self.elementwise_affine}'


def get_dropout_layer(p):
    return nn.Dropout(p, inplace=True) if p > 0 else nn.Identity()


class FFN(nn.Module):
    def __init__(self, in_features, hidden_features=None, out_features=None, drop=0., fused_mlp=False):
        super().__init__()
        self.fused_mlp_func = fused_mlp_func if fused_mlp else None
        out_features = out_features or in_features
        hidden_features = hidden_features or in_features
        self.fc1 = nn.Linear(in_features, hidden_features)
        self.act = nn.GELU(approximate='tanh')
        self.fc2 = nn.Linear(hidden_features, out_features)
        self.drop = get_dropout_layer(drop)
        self.heuristic = -1
    
    def forward(self, x):
        if self.fused_mlp_func is not None:
            return self.drop(self.fused_mlp_func(
                x=x,
                weight1=self.fc1.weight,
                weight2=self.fc2.weight,
                bias1=self.fc1.bias,
                bias2=self.fc2.bias,
                activation='gelu_approx',
                save_pre_act=self.training,
                return_residual=False,
                checkpoint_lvl=0,
                heuristic=self.heuristic,
                process_group=None,
            ))
        else:
            return self.drop(self.fc2(self.act(self.fc1(x))))
    
    def extra_repr(self) -> str:
        return f'fused_mlp={self.fused_mlp_func is not None}'

class Qwen3MLP(nn.Module):
    def __init__(self, hidden_size, intermediate_size):
        super().__init__()
        self.hidden_size = hidden_size
        self.intermediate_size = intermediate_size
        self.gate_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.up_proj = nn.Linear(self.hidden_size, self.intermediate_size, bias=False)
        self.down_proj = nn.Linear(self.intermediate_size, self.hidden_size, bias=False)
        self.act_fn = nn.SiLU()

    def forward(self, x):
        down_proj = self.down_proj(self.act_fn(self.gate_proj(x)) * self.up_proj(x))
        return down_proj

class FFNSwiGLU(nn.Module):
    def __init__(self, in_features, hidden_features, out_features=None, drop=0., fused_mlp=False):
        super().__init__()
        self.fused_mlp_func = None
        hidden_features = round(2 * hidden_features / 3 / 256) * 256
        
        out_features = out_features or in_features
        self.fcg = nn.Linear(in_features, hidden_features, bias=False)
        self.fc1 = nn.Linear(in_features, hidden_features, bias=False)
        self.fc2 = nn.Linear(hidden_features, out_features, bias=False)
        self.drop = get_dropout_layer(drop)
    
    def forward(self, x):
        return self.drop(self.fc2( F.silu(self.fcg(x), inplace=True).mul_(self.fc1(x)) ))
    
    def extra_repr(self) -> str:
        return f'fused_mlp={self.fused_mlp_func is not None}'

def repeat_kv(hidden_states: torch.Tensor, n_rep: int) -> torch.Tensor:
    """
    This is the equivalent of torch.repeat_interleave(x, dim=1, repeats=n_rep). The hidden states go from (batch,
    num_key_value_heads, seqlen, head_dim) to (batch, num_attention_heads, seqlen, head_dim)
    """
    batch, num_key_value_heads, slen, head_dim = hidden_states.shape
    if n_rep == 1:
        return hidden_states
    hidden_states = hidden_states[:, :, None, :, :].expand(batch, num_key_value_heads, n_rep, slen, head_dim)
    return hidden_states.reshape(batch, num_key_value_heads * n_rep, slen, head_dim)

class SelfAttention(nn.Module):
    def __init__(
        self, embed_dim=768, num_heads=12, num_key_value_heads=-1,
        use_flex_attn=False, 
        pad_to_multiplier=1, rope2d_normalized_by_hw=0,
        mask_type='var', context_frames=1000000, steps_per_frame=4,
        arch='var',
        qwen_qkvo_bias=False,
    ):
        """
        :param embed_dim: model's width
        :param num_heads: num heads of multi-head attention
        """
        super().__init__()
        assert embed_dim % num_heads == 0
        assert num_key_value_heads == -1 or num_heads % num_key_value_heads == 0
        
        self.embed_dim = embed_dim
        self.num_heads, self.head_dim = num_heads, embed_dim // num_heads
        self.num_key_value_heads = num_key_value_heads if num_key_value_heads > 0 else num_heads
        self.arch = arch
        if self.arch == 'qwen':
            self.q_proj = nn.Linear(embed_dim, self.num_heads*self.head_dim, bias=qwen_qkvo_bias)
            self.k_proj = nn.Linear(embed_dim, self.num_key_value_heads*self.head_dim, bias=qwen_qkvo_bias)
            self.v_proj = nn.Linear(embed_dim, self.num_key_value_heads*self.head_dim, bias=qwen_qkvo_bias)
            self.o_proj = nn.Linear(self.num_heads*self.head_dim, embed_dim, bias=qwen_qkvo_bias)
            self.q_norm = FastRMSNorm(self.head_dim)
            self.k_norm = FastRMSNorm(self.head_dim)
            self.num_key_value_groups = self.num_heads // self.num_key_value_heads
        else:
            raise ValueError(f'arch {self.arch} not supported')
        
        self.caching = False    # kv caching: only used during inference
        self.cached_k = {}    # kv caching: only used during inference
        self.cached_v = {}    # kv caching: only used during inference

        self.use_flex_attn = use_flex_attn
        self.pad_to_multiplier = pad_to_multiplier

        self.rope2d_normalized_by_hw = rope2d_normalized_by_hw
        self.mask_type = mask_type
        self.context_frames = context_frames
        self.steps_per_frame = steps_per_frame
    
    def kv_caching(self, enable: bool): # kv caching: only used during inference
        self.caching = enable
        self.cached_k = {}
        self.cached_v = {}

    # NOTE: attn_bias_or_two_vector is None during inference
    def forward(self, x, attn_bias_or_two_vector: Union[torch.Tensor, Tuple[torch.IntTensor, torch.IntTensor]], attn_fn=None, rope2d_freqs_grid=[], scale_schedule=[], scale_ind=0, context_info=None, last_repetition_step=True, ref_text_scale_inds=[]):
        """
        :param (fp32) x: shaped (B or batch_size, L or seq_length, C or hidden_dim); if seq-parallel is used, the `L` dim would be sharded (L = raw_seq_len//sp_size)
        :param (fp32) attn_bias_or_two_vector:
                if not using_flash:
                    a block-wise, lower-triangle matrix, like:
                    [[[[0, -, -, -, -, -, -, -, -, -, -, -, -, -],
                    [0, 0, 0, 0, 0, -, -, -, -, -, -, -, -, -],
                    [0, 0, 0, 0, 0, -, -, -, -, -, -, -, -, -],
                    [0, 0, 0, 0, 0, -, -, -, -, -, -, -, -, -],
                    [0, 0, 0, 0, 0, -, -, -, -, -, -, -, -, -],
                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0],
                    [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]]]]
                    where 0 means visible and - means invisible (-inf)
                else:
                    a tuple of two 1-dim int vector (VAR_visible_kvlen, VAR_invisible_qlen)
        :return: shaped (B or batch_size, L or seq_length, C or hidden_dim); if seq-parallel is used, the `L` dim would be sharded
        """
        # x: fp32
        B, L, C = x.shape

        if self.arch == 'qwen':
            hidden_states = x
            input_shape = hidden_states.shape[:-1]
            hidden_shape = (*input_shape, -1, self.head_dim)

            query_states = self.q_norm(self.q_proj(hidden_states).view(hidden_shape)).transpose(1, 2)
            key_states = self.k_norm(self.k_proj(hidden_states).view(hidden_shape)).transpose(1, 2) # batch, num_key_value_heads, slen, head_dim
            value_states = self.v_proj(hidden_states).view(hidden_shape).transpose(1, 2) # batch, num_key_value_heads, slen, head_dim

            if sp_manager.sp_on():
                # Headnum need to be sharded and L needs to be gathered
                # [B, H, raw_L/sp, C] --> [B, H/sp, raw_L, C]
                sdim = 1
                gdim = 2
                L = L * sp_manager.get_sp_size()
                C = C // sp_manager.get_sp_size()
                query_states = sp_all_to_all(query_states, sdim, gdim)
                key_states = sp_all_to_all(key_states, sdim, gdim)
                value_states = sp_all_to_all(value_states, sdim, gdim)

            query_states, key_states = apply_rotary_emb(query_states, key_states, rope2d_freqs_grid)
            if self.caching:    # kv caching: only used during inference
                if last_repetition_step:
                    self.cached_k[scale_ind] = key_states
                    self.cached_v[scale_ind] = value_states
                if isinstance(scale_ind, int):
                    ref_scale_inds = context_info[scale_ind]['ref_sids'] + ref_text_scale_inds
                    key_states = torch.cat([self.cached_k[ind] for ind in ref_scale_inds] + [key_states], dim=2)
                    value_states = torch.cat([self.cached_v[ind] for ind in ref_scale_inds] + [value_states], dim=2)
                
                    ref_scale_2_last_use_scale = [-1 for _ in range(len(context_info))]
                    for si in range(len(context_info)):
                        for ref_si in context_info[si]['ref_sids']:
                            ref_scale_2_last_use_scale[ref_si] = si
                    for ref_si in range(scale_ind):
                        if (ref_scale_2_last_use_scale[ref_si] < scale_ind) and (self.cached_k[ref_si] is not None):
                            tmpk, tmpv = self.cached_k[ref_si], self.cached_v[ref_si]
                            self.cached_k[ref_si], self.cached_v[ref_si] = None, None
                            del tmpk, tmpv

            key_states = repeat_kv(key_states, self.num_key_value_groups)
            value_states = repeat_kv(value_states, self.num_key_value_groups)
            scale = self.head_dim**-0.5
            if self.use_flex_attn and attn_fn is not None:
                attn_output = attn_fn(query_states.to(value_states.dtype), key_states.to(value_states.dtype), value_states, scale=scale).transpose(1, 2).reshape(B, L, C)
            else:
                # fa2, flash_attn_func input/output should be (batch_size, seqlen, nheads, headdim)
                from flash_attn import flash_attn_qkvpacked_func, flash_attn_func
                attn_output = flash_attn_func(query_states.permute([0,2,1,3]).to(torch.bfloat16), key_states.permute([0,2,1,3]).to(torch.bfloat16), value_states.permute([0,2,1,3]).to(torch.bfloat16), softmax_scale=scale)
                attn_output = attn_output.reshape(B, L, C)

                # fa3, flash_attn_func input/output should be (batch_size, seqlen, nheads, headdim)
                # from flash_attn_interface import flash_attn_qkvpacked_func, flash_attn_func
                # attn_output = flash_attn_func(query_states.permute([0,2,1,3]).to(torch.bfloat16), key_states.permute([0,2,1,3]).to(torch.bfloat16), value_states.permute([0,2,1,3]).to(torch.bfloat16), softmax_scale=scale)
                # attn_output = attn_output[0].reshape(B, L, C)
                
                # slow attn
                # attn_output = slow_attn(query=query_states, key=key_states, value=value_states, scale=scale, attn_mask=attn_bias_or_two_vector, dropout_p=0).transpose(1, 2).reshape(B, L, C)
            if sp_manager.sp_on():
                # [B, raw_L, C/sp] --> [B, raw_L/sp, C]
                sdim = 1
                gdim = 2
                attn_output = sp_all_to_all(attn_output, sdim, gdim)

            attn_output = self.o_proj(attn_output)

            return attn_output
        
        # qkv: amp, bf16
        qkv = F.linear(input=x, weight=self.mat_qkv.weight, bias=torch.cat((self.q_bias, self.zero_k_bias, self.v_bias))).view(B, L, 3, self.num_heads, self.head_dim)  # BL3Hc
        q, k, v = qkv.permute(2, 0, 3, 1, 4).unbind(dim=0); L_dim = 2   # q or k or v: all are shaped in (B:batch_size, H:heads, L:seq_len, c:head_dim), this way
        
        scale_mul = self.scale_mul_1H11.clamp_max(self.max_scale_mul).exp() # 11H1 (flash), or 1H11 (not flash)
        q = F.normalize(q, dim=-1, eps=1e-12).mul(scale_mul).contiguous()   # fp32
        k = F.normalize(k, dim=-1, eps=1e-12).contiguous()                  # fp32
        v = v.contiguous()                                                  # bf16

        if sp_manager.sp_on():
            # Headnum need to be sharded and L needs to be gathered
            # [B, H, raw_L/sp, C] --> [B, H/sp, raw_L, C]
            sdim = 1
            gdim = 2

            L = L * sp_manager.get_sp_size()
            C = C // sp_manager.get_sp_size()

            q = sp_all_to_all(q, sdim, gdim)
            k = sp_all_to_all(k, sdim, gdim)
            v = sp_all_to_all(v, sdim, gdim)


        q, k = apply_rotary_emb(q, k, rope2d_freqs_grid) #, freqs_cis=freqs_cis)
        if self.caching:    # kv caching: only used during inference
            if last_repetition_step:
                self.cached_k.append(k)
                self.cached_v.append(v)
            if scale_ind >= 0:
                ref_scale_inds = context_info[scale_ind]['ref_sids']
                k = torch.cat([self.cached_k[0]] + [self.cached_k[ind+1] for ind in ref_scale_inds] + [k], dim=L_dim)
                v = torch.cat([self.cached_v[0]] + [self.cached_v[ind+1] for ind in ref_scale_inds] + [v], dim=L_dim)

            ref_scale_2_last_use_scale = [-1 for _ in range(len(context_info))]
            for si in range(len(context_info)):
                for ref_si in context_info[si]['ref_sids']:
                    ref_scale_2_last_use_scale[ref_si] = si
            for ref_si in range(scale_ind):
                if (ref_scale_2_last_use_scale[ref_si] < scale_ind) and (self.cached_k[ref_si+1] is not None):
                    tmpk, tmpv = self.cached_k[ref_si+1], self.cached_v[ref_si+1]
                    self.cached_k[ref_si+1], self.cached_v[ref_si+1] = None, None
                    del tmpk, tmpv
        
        # if self.cos_attn: q, k are in fp32; v is in bf16
        # else: q, k, v are in bf16
        if self.use_flex_attn and attn_fn is not None:
            oup = attn_fn(q.to(v.dtype), k.to(v.dtype), v, scale=self.scale).transpose(1, 2).reshape(B, L, C)
        else:
            # oup = slow_attn(query=q, key=k, value=v, scale=self.scale, attn_mask=attn_bias_or_two_vector, dropout_p=0).transpose(1, 2).reshape(B, L, C)
            # fa2, flash_attn_func input/output should be (batch_size, seqlen, nheads, headdim)
            from flash_attn import flash_attn_qkvpacked_func, flash_attn_func
            oup = flash_attn_func(q.permute([0,2,1,3]).to(torch.bfloat16), k.permute([0,2,1,3]).to(torch.bfloat16), v.permute([0,2,1,3]).to(torch.bfloat16), softmax_scale=self.scale)
            oup = oup.reshape(B, L, C)
        # oup: bf16

        if sp_manager.sp_on():
            # [B, raw_L, C/sp] --> [B, raw_L/sp, C]
            sdim = 1
            gdim = 2
            oup = sp_all_to_all(oup, sdim, gdim)

        return self.proj_drop(self.proj(oup))
    
class SelfAttnBlock(nn.Module):
    def __init__(
        self,
        embed_dim,
        cond_dim,
        num_heads,
        num_key_value_heads,
        mlp_ratio=4.0,
        use_flex_attn=False,
        pad_to_multiplier=1,
        rope2d_normalized_by_hw=False,
        mask_type="",
        context_frames=-1,
        steps_per_frame=-1,
        arch="var",
        qwen_qkvo_bias=False,
        inject_sync=False,
    ):
        super(SelfAttnBlock, self).__init__()
        self.C, self.D = embed_dim, cond_dim
        self.arch=arch
        self.attn = SelfAttention(
            embed_dim=embed_dim, num_heads=num_heads, num_key_value_heads=num_key_value_heads,
            use_flex_attn=use_flex_attn, pad_to_multiplier=pad_to_multiplier, rope2d_normalized_by_hw=rope2d_normalized_by_hw,
            mask_type=mask_type, context_frames=context_frames, steps_per_frame=steps_per_frame, arch=arch, qwen_qkvo_bias=qwen_qkvo_bias,
        )
        if self.arch == 'qwen':
            self.mlp = Qwen3MLP(hidden_size=embed_dim, intermediate_size=round(embed_dim * mlp_ratio / 256) * 256)
            self.input_layernorm = FastRMSNorm(embed_dim)
            self.post_attention_layernorm = FastRMSNorm(embed_dim)
            self.inject_sync = inject_sync
        else:
            raise ValueError(f'arch {self.arch} not supported')
        
    # NOTE: attn_bias_or_two_vector is None during inference
    def forward(self, x, cond_BD, ca_kv, attn_bias_or_two_vector, attn_fn=None, rope2d_freqs_grid=[], scale_schedule=[], scale_ind=0, context_info=None, last_repetition_step=True, ref_text_scale_inds=[]):
        residual = x
        hidden_states = x
        hidden_states = self.input_layernorm(hidden_states)
        hidden_states = self.attn(hidden_states, attn_bias_or_two_vector, attn_fn, rope2d_freqs_grid, scale_schedule, scale_ind, context_info, last_repetition_step, ref_text_scale_inds)
        hidden_states = residual + hidden_states
        # Fully Connected
        residual = hidden_states
        hidden_states = self.post_attention_layernorm(hidden_states)
        hidden_states = self.mlp(hidden_states)
        hidden_states = residual + hidden_states
        return hidden_states
    

if __name__ == '__main__':
    pass