File size: 12,217 Bytes
50eec37
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
import torch.nn as nn
from dataclasses import dataclass
from typing import Optional, Any

from comfy.ldm.modules.attention import optimized_attention_for_device
import comfy.model_management
import comfy.ldm.common_dit

import comfy.model_management

@dataclass
class Llama2Config:
    vocab_size: int = 128320
    hidden_size: int = 4096
    intermediate_size: int = 14336
    num_hidden_layers: int = 32
    num_attention_heads: int = 32
    num_key_value_heads: int = 8
    max_position_embeddings: int = 8192
    rms_norm_eps: float = 1e-5
    rope_theta: float = 500000.0
    transformer_type: str = "llama"
    head_dim = 128
    rms_norm_add = False
    mlp_activation = "silu"

@dataclass
class Gemma2_2B_Config:
    vocab_size: int = 256000
    hidden_size: int = 2304
    intermediate_size: int = 9216
    num_hidden_layers: int = 26
    num_attention_heads: int = 8
    num_key_value_heads: int = 4
    max_position_embeddings: int = 8192
    rms_norm_eps: float = 1e-6
    rope_theta: float = 10000.0
    transformer_type: str = "gemma2"
    head_dim = 256
    rms_norm_add = True
    mlp_activation = "gelu_pytorch_tanh"

class RMSNorm(nn.Module):
    def __init__(self, dim: int, eps: float = 1e-5, add=False, device=None, dtype=None):
        super().__init__()
        self.eps = eps
        self.weight = nn.Parameter(torch.empty(dim, device=device, dtype=dtype))
        self.add = add

    def forward(self, x: torch.Tensor):
        w = self.weight
        if self.add:
            w = w + 1.0

        return comfy.ldm.common_dit.rms_norm(x, w, self.eps)



def rotate_half(x):
    """Rotates half the hidden dims of the input."""
    x1 = x[..., : x.shape[-1] // 2]
    x2 = x[..., x.shape[-1] // 2 :]
    return torch.cat((-x2, x1), dim=-1)


def precompute_freqs_cis(head_dim, seq_len, theta, device=None):
    theta_numerator = torch.arange(0, head_dim, 2, device=device).float()
    inv_freq = 1.0 / (theta ** (theta_numerator / head_dim))

    position_ids = torch.arange(0, seq_len, device=device).unsqueeze(0)

    inv_freq_expanded = inv_freq[None, :, None].float().expand(position_ids.shape[0], -1, 1)
    position_ids_expanded = position_ids[:, None, :].float()
    freqs = (inv_freq_expanded.float() @ position_ids_expanded.float()).transpose(1, 2)
    emb = torch.cat((freqs, freqs), dim=-1)
    cos = emb.cos()
    sin = emb.sin()
    return (cos, sin)


def apply_rope(xq, xk, freqs_cis):
    cos = freqs_cis[0].unsqueeze(1)
    sin = freqs_cis[1].unsqueeze(1)
    q_embed = (xq * cos) + (rotate_half(xq) * sin)
    k_embed = (xk * cos) + (rotate_half(xk) * sin)
    return q_embed, k_embed


class Attention(nn.Module):
    def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None):
        super().__init__()
        self.num_heads = config.num_attention_heads
        self.num_kv_heads = config.num_key_value_heads
        self.hidden_size = config.hidden_size

        self.head_dim = config.head_dim
        self.inner_size = self.num_heads * self.head_dim

        ops = ops or nn
        self.q_proj = ops.Linear(config.hidden_size, self.inner_size, bias=False, device=device, dtype=dtype)
        self.k_proj = ops.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False, device=device, dtype=dtype)
        self.v_proj = ops.Linear(config.hidden_size, self.num_kv_heads * self.head_dim, bias=False, device=device, dtype=dtype)
        self.o_proj = ops.Linear(self.inner_size, config.hidden_size, bias=False, device=device, dtype=dtype)

    def forward(
        self,
        hidden_states: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        freqs_cis: Optional[torch.Tensor] = None,
        optimized_attention=None,
    ):
        batch_size, seq_length, _ = hidden_states.shape
        xq = self.q_proj(hidden_states)
        xk = self.k_proj(hidden_states)
        xv = self.v_proj(hidden_states)

        xq = xq.view(batch_size, seq_length, self.num_heads, self.head_dim).transpose(1, 2)
        xk = xk.view(batch_size, seq_length, self.num_kv_heads, self.head_dim).transpose(1, 2)
        xv = xv.view(batch_size, seq_length, self.num_kv_heads, self.head_dim).transpose(1, 2)

        xq, xk = apply_rope(xq, xk, freqs_cis=freqs_cis)

        xk = xk.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1)
        xv = xv.repeat_interleave(self.num_heads // self.num_kv_heads, dim=1)

        output = optimized_attention(xq, xk, xv, self.num_heads, mask=attention_mask, skip_reshape=True)
        return self.o_proj(output)

class MLP(nn.Module):
    def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None):
        super().__init__()
        ops = ops or nn
        self.gate_proj = ops.Linear(config.hidden_size, config.intermediate_size, bias=False, device=device, dtype=dtype)
        self.up_proj = ops.Linear(config.hidden_size, config.intermediate_size, bias=False, device=device, dtype=dtype)
        self.down_proj = ops.Linear(config.intermediate_size, config.hidden_size, bias=False, device=device, dtype=dtype)
        if config.mlp_activation == "silu":
            self.activation = torch.nn.functional.silu
        elif config.mlp_activation == "gelu_pytorch_tanh":
            self.activation = lambda a: torch.nn.functional.gelu(a, approximate="tanh")

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

class TransformerBlock(nn.Module):
    def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None):
        super().__init__()
        self.self_attn = Attention(config, device=device, dtype=dtype, ops=ops)
        self.mlp = MLP(config, device=device, dtype=dtype, ops=ops)
        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, device=device, dtype=dtype)
        self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, device=device, dtype=dtype)

    def forward(
        self,
        x: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        freqs_cis: Optional[torch.Tensor] = None,
        optimized_attention=None,
    ):
        # Self Attention
        residual = x
        x = self.input_layernorm(x)
        x = self.self_attn(
            hidden_states=x,
            attention_mask=attention_mask,
            freqs_cis=freqs_cis,
            optimized_attention=optimized_attention,
        )
        x = residual + x

        # MLP
        residual = x
        x = self.post_attention_layernorm(x)
        x = self.mlp(x)
        x = residual + x

        return x

class TransformerBlockGemma2(nn.Module):
    def __init__(self, config: Llama2Config, device=None, dtype=None, ops: Any = None):
        super().__init__()
        self.self_attn = Attention(config, device=device, dtype=dtype, ops=ops)
        self.mlp = MLP(config, device=device, dtype=dtype, ops=ops)
        self.input_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
        self.post_attention_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
        self.pre_feedforward_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
        self.post_feedforward_layernorm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)

    def forward(
        self,
        x: torch.Tensor,
        attention_mask: Optional[torch.Tensor] = None,
        freqs_cis: Optional[torch.Tensor] = None,
        optimized_attention=None,
    ):
        # Self Attention
        residual = x
        x = self.input_layernorm(x)
        x = self.self_attn(
            hidden_states=x,
            attention_mask=attention_mask,
            freqs_cis=freqs_cis,
            optimized_attention=optimized_attention,
        )

        x = self.post_attention_layernorm(x)
        x = residual + x

        # MLP
        residual = x
        x = self.pre_feedforward_layernorm(x)
        x = self.mlp(x)
        x = self.post_feedforward_layernorm(x)
        x = residual + x

        return x

class Llama2_(nn.Module):
    def __init__(self, config, device=None, dtype=None, ops=None):
        super().__init__()
        self.config = config
        self.vocab_size = config.vocab_size

        self.embed_tokens = ops.Embedding(
            config.vocab_size,
            config.hidden_size,
            device=device,
            dtype=dtype
        )
        if self.config.transformer_type == "gemma2":
            transformer = TransformerBlockGemma2
            self.normalize_in = True
        else:
            transformer = TransformerBlock
            self.normalize_in = False

        self.layers = nn.ModuleList([
            transformer(config, device=device, dtype=dtype, ops=ops)
            for _ in range(config.num_hidden_layers)
        ])
        self.norm = RMSNorm(config.hidden_size, eps=config.rms_norm_eps, add=config.rms_norm_add, device=device, dtype=dtype)
        # self.lm_head = ops.Linear(config.hidden_size, config.vocab_size, bias=False, device=device, dtype=dtype)

    def forward(self, x, attention_mask=None, embeds=None, num_tokens=None, intermediate_output=None, final_layer_norm_intermediate=True, dtype=None):
        if embeds is not None:
            x = embeds
        else:
            x = self.embed_tokens(x, out_dtype=dtype)

        if self.normalize_in:
            x *= self.config.hidden_size ** 0.5

        freqs_cis = precompute_freqs_cis(self.config.head_dim,
                                         x.shape[1],
                                         self.config.rope_theta,
                                         device=x.device)

        mask = None
        if attention_mask is not None:
            mask = 1.0 - attention_mask.to(x.dtype).reshape((attention_mask.shape[0], 1, -1, attention_mask.shape[-1])).expand(attention_mask.shape[0], 1, attention_mask.shape[-1], attention_mask.shape[-1])
            mask = mask.masked_fill(mask.to(torch.bool), float("-inf"))

        causal_mask = torch.empty(x.shape[1], x.shape[1], dtype=x.dtype, device=x.device).fill_(float("-inf")).triu_(1)
        if mask is not None:
            mask += causal_mask
        else:
            mask = causal_mask
        optimized_attention = optimized_attention_for_device(x.device, mask=mask is not None, small_input=True)

        intermediate = None
        if intermediate_output is not None:
            if intermediate_output < 0:
                intermediate_output = len(self.layers) + intermediate_output

        for i, layer in enumerate(self.layers):
            x = layer(
                x=x,
                attention_mask=mask,
                freqs_cis=freqs_cis,
                optimized_attention=optimized_attention,
            )
            if i == intermediate_output:
                intermediate = x.clone()

        x = self.norm(x)
        if intermediate is not None and final_layer_norm_intermediate:
            intermediate = self.norm(intermediate)

        return x, intermediate

class BaseLlama:
    def get_input_embeddings(self):
        return self.model.embed_tokens

    def set_input_embeddings(self, embeddings):
        self.model.embed_tokens = embeddings

    def forward(self, input_ids, *args, **kwargs):
        return self.model(input_ids, *args, **kwargs)


class Llama2(BaseLlama, torch.nn.Module):
    def __init__(self, config_dict, dtype, device, operations):
        super().__init__()
        config = Llama2Config(**config_dict)
        self.num_layers = config.num_hidden_layers

        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
        self.dtype = dtype


class Gemma2_2B(BaseLlama, torch.nn.Module):
    def __init__(self, config_dict, dtype, device, operations):
        super().__init__()
        config = Gemma2_2B_Config(**config_dict)
        self.num_layers = config.num_hidden_layers

        self.model = Llama2_(config, device=device, dtype=dtype, ops=operations)
        self.dtype = dtype