File size: 11,515 Bytes
3e936b2
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.distributed as dist
from wan.modules.attention import attention
from wan.modules.model import WanRMSNorm

class MemoryCrossAttentionLayer(nn.Module):

    def __init__(self, dim, num_heads, ffn_dim, qk_norm=True, eps=1e-06):
        super().__init__()
        self.num_heads = num_heads
        self.head_dim = dim // num_heads
        self.q = nn.Linear(dim, dim)
        self.o = nn.Linear(dim, dim)
        self.norm_q = WanRMSNorm(dim, eps=eps) if qk_norm else nn.Identity()
        self.norm1 = WanRMSNorm(dim, eps=eps)
        self.norm2 = WanRMSNorm(dim, eps=eps)
        self.ffn = nn.Sequential(nn.Linear(dim, ffn_dim), nn.GELU(approximate='tanh'), nn.Linear(ffn_dim, dim))

    def forward(self, query_state, cached_k, cached_v):
        b = query_state.shape[0]
        m = query_state.shape[1]
        n, d = (self.num_heads, self.head_dim)
        residual = query_state
        h = self.norm1(query_state)
        q = self.norm_q(self.q(h)).view(b, m, n, d)
        out = attention(q, cached_k, cached_v)
        out = out.flatten(2).type_as(h)
        out = self.o(out)
        query_state = residual + out
        residual = query_state
        query_state = residual + self.ffn(self.norm2(query_state))
        return query_state

class QueryMemoryEncoder(nn.Module):

    def __init__(self, config):
        super().__init__()
        Q_frames = getattr(config, 'Q_frames', 3)
        tokens_per_frame = getattr(config, 'tokens_per_frame', 1560)
        n_encoder_layers = getattr(config, 'n_encoder_layers', 2)
        hidden_dim = getattr(config, 'hidden_dim', 1536)
        num_heads = getattr(config, 'num_heads', 12)
        head_dim = getattr(config, 'head_dim', 128)
        ffn_dim = hidden_dim * 4
        gate_init_bias = getattr(config, 'gate_init_bias', 2.0)
        qk_norm = getattr(config, 'qk_norm', True)
        eps = 1e-06
        self.M = Q_frames * tokens_per_frame
        self.hidden_dim = hidden_dim
        self.num_heads = num_heads
        self.head_dim = head_dim
        self.Q_frames = Q_frames
        self.tokens_per_frame = tokens_per_frame
        self.use_batch_update = getattr(config, 'use_batch_update', False)
        self.batch_update_interval = getattr(config, 'batch_update_interval', 1)
        self.use_sink_anchor = getattr(config, 'use_sink_anchor', False)
        self.use_vib = getattr(config, 'use_vib', False)
        self.bptt_clips = getattr(config, 'bptt_clips', 1)
        self.encoder_lr_multiplier = getattr(config, 'encoder_lr_multiplier', 5.0)
        self.normalize_memory_k = getattr(config, 'normalize_memory_k', False)
        self.use_residual_update = getattr(config, 'use_residual_update', False)
        self.use_post_norm = getattr(config, 'use_post_norm', False)
        self.memory_recache = getattr(config, 'memory_recache', False)
        self.num_query_groups = getattr(config, 'num_query_groups', 1)
        initializer_range = getattr(config, 'initializer_range', 0.014)
        self.layers = nn.ModuleList([MemoryCrossAttentionLayer(hidden_dim, num_heads, ffn_dim, qk_norm, eps) for _ in range(n_encoder_layers)])
        if self.use_post_norm:
            self.post_norm = WanRMSNorm(hidden_dim, eps=eps)
        if self.num_query_groups > 1:
            self.query_inits = nn.ParameterList([nn.Parameter(torch.randn(1, self.M, hidden_dim) * initializer_range) for _ in range(self.num_query_groups)])
            self.to_k_groups = nn.ModuleList([nn.Linear(hidden_dim, num_heads * head_dim) for _ in range(self.num_query_groups)])
            self.to_v_groups = nn.ModuleList([nn.Linear(hidden_dim, num_heads * head_dim) for _ in range(self.num_query_groups)])
            if self.normalize_memory_k:
                self.norm_k_out_groups = nn.ModuleList([WanRMSNorm(num_heads * head_dim, eps=eps) for _ in range(self.num_query_groups)])
            self.connector_projs = nn.ModuleList([nn.Sequential(nn.Linear(hidden_dim, hidden_dim), nn.GELU(approximate='tanh'), nn.Linear(hidden_dim, hidden_dim), WanRMSNorm(hidden_dim, eps=eps)) for _ in range(self.num_query_groups)])
            self.gate_linears = nn.ModuleList([nn.Linear(hidden_dim * 2, hidden_dim) for _ in range(self.num_query_groups)])
            for gl in self.gate_linears:
                with torch.no_grad():
                    gl.bias.fill_(gate_init_bias)
        else:
            self.query_init = nn.Parameter(torch.randn(1, self.M, hidden_dim) * initializer_range)
            self.connector_proj = nn.Sequential(nn.Linear(hidden_dim, hidden_dim), nn.GELU(approximate='tanh'), nn.Linear(hidden_dim, hidden_dim), WanRMSNorm(hidden_dim, eps=eps))
            self.gate_linear = nn.Linear(hidden_dim * 2, hidden_dim)
            with torch.no_grad():
                self.gate_linear.bias.fill_(gate_init_bias)
            self.to_k = nn.Linear(hidden_dim, num_heads * head_dim)
            self.to_v = nn.Linear(hidden_dim, num_heads * head_dim)
            if self.normalize_memory_k:
                self.norm_k_out = WanRMSNorm(num_heads * head_dim, eps=eps)
        if self.use_vib:
            self.mu_proj = nn.Linear(hidden_dim, hidden_dim)
            self.logvar_proj = nn.Linear(hidden_dim, hidden_dim)
        self._evicted_k_buffer = []
        self._evicted_v_buffer = []
        self.query_state = None
        self.has_history = False
        self._update_count = 0
        total_params = sum((p.numel() for p in self.parameters()))

    def reset(self, batch_size=1, device=None, dtype=None):

        def _expand_init(init):
            if init.dim() != 3 or init.shape[1] != self.M:
                raise RuntimeError(f'query_init has unexpected shape {list(init.shape)} (expected [1, {self.M}, {self.hidden_dim}]). This usually means the encoder was placed inside FSDP, which flattens parameters. The encoder must remain outside FSDP (attached via object.__setattr__).')
            state = init.expand(batch_size, -1, -1).clone()
            if device is not None:
                state = state.to(device=device)
            if dtype is not None:
                state = state.to(dtype=dtype)
            return state
        if self.num_query_groups > 1:
            self.query_states = [_expand_init(init) for init in self.query_inits]
            self.query_state = self.query_states[0]
        else:
            self.query_state = _expand_init(self.query_init)
        self.has_history = False
        self._update_count = 0
        self._evicted_k_buffer = []
        self._evicted_v_buffer = []
        shape = list(self.query_states[0].shape) if self.num_query_groups > 1 else list(self.query_state.shape)

    def update(self, evicted_k, evicted_v, sink_k=None, sink_v=None):
        if self.query_state is None:
            self.reset(batch_size=evicted_k.shape[0], device=evicted_k.device, dtype=evicted_k.dtype if evicted_k.dtype.is_floating_point else torch.bfloat16)
        query_state = self.query_state
        if self.use_batch_update:
            self._evicted_k_buffer.append(evicted_k)
            self._evicted_v_buffer.append(evicted_v)
            if len(self._evicted_k_buffer) < self.batch_update_interval:
                return 0.0
            evicted_k = torch.cat(self._evicted_k_buffer, dim=1)
            evicted_v = torch.cat(self._evicted_v_buffer, dim=1)
            self._evicted_k_buffer = []
            self._evicted_v_buffer = []
        if self.use_sink_anchor and sink_k is not None:
            ctx_k = torch.cat([sink_k, evicted_k], dim=1)
            ctx_v = torch.cat([sink_v, evicted_v], dim=1)
        else:
            ctx_k = evicted_k
            ctx_v = evicted_v

        def _update_single_group(qs, connector_proj, gate_linear):
            state = qs
            for layer in self.layers:
                state = layer(state, ctx_k, ctx_v)
            if self.use_residual_update:
                new_state = state
                gate_mean = None
            else:
                projected = connector_proj(state)
                gate = torch.sigmoid(gate_linear(torch.cat([qs, projected], dim=-1)))
                new_state = gate * qs + (1 - gate) * projected
                gate_mean = gate.mean().item()
            if self.use_post_norm:
                new_state = self.post_norm(new_state)
            return (new_state, gate_mean)
        kl_loss = 0.0
        if self.num_query_groups > 1:
            gate_means = []
            for g in range(self.num_query_groups):
                new_state, gm = _update_single_group(self.query_states[g], self.connector_projs[g], self.gate_linears[g])
                if self.use_vib and self.training:
                    mu = self.mu_proj(new_state)
                    logvar = self.logvar_proj(new_state)
                    std = torch.exp(0.5 * logvar)
                    new_state = mu + torch.randn_like(std) * std
                    kl_loss += -0.5 * (1 + logvar - mu.pow(2) - logvar.exp()).sum()
                elif self.use_vib:
                    new_state = self.mu_proj(new_state)
                self.query_states[g] = new_state
                gate_means.append(gm)
            self.query_state = self.query_states[0]
            gate_str = ', '.join((f"g{i}={('res' if gm is None else f'{gm:.3f}')}" for i, gm in enumerate(gate_means)))
            norm_str = ', '.join((f'g{i}={self.query_states[i].norm().item():.1f}' for i in range(self.num_query_groups)))
        else:
            new_state, gate_mean = _update_single_group(query_state, self.connector_proj, self.gate_linear)
            if self.use_vib and self.training:
                mu = self.mu_proj(new_state)
                logvar = self.logvar_proj(new_state)
                std = torch.exp(0.5 * logvar)
                new_state = mu + torch.randn_like(std) * std
                kl_loss = -0.5 * (1 + logvar - mu.pow(2) - logvar.exp()).sum()
            elif self.use_vib:
                new_state = self.mu_proj(new_state)
            self.query_state = new_state
            gate_str = f"{('residual_update' if gate_mean is None else f'gate_mean={gate_mean:.3f}')}"
            norm_str = f'{new_state.norm().item():.1f}'
        self.has_history = True
        self._update_count += 1
        return kl_loss

    def get_kv(self, group_index=None):
        if not self.has_history:
            return None
        if self.num_query_groups > 1 and group_index is not None:
            state = self.query_states[group_index]
            k = self.to_k_groups[group_index](state)
            if self.normalize_memory_k:
                k = self.norm_k_out_groups[group_index](k)
            k = k.view(state.shape[0], self.M, self.num_heads, self.head_dim)
            v = self.to_v_groups[group_index](state).view(state.shape[0], self.M, self.num_heads, self.head_dim)
            return (k, v)
        else:
            state = self.query_state
            k = self.to_k(state)
            if self.normalize_memory_k:
                k = self.norm_k_out(k)
            k = k.view(state.shape[0], self.M, self.num_heads, self.head_dim)
            v = self.to_v(state).view(state.shape[0], self.M, self.num_heads, self.head_dim)
            return (k, v)

    def detach_state(self):
        if self.num_query_groups > 1:
            self.query_states = [s.detach() for s in self.query_states]
            self.query_state = self.query_states[0]
        elif self.query_state is not None:
            self.query_state = self.query_state.detach()