File size: 7,702 Bytes
aa16b75
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
from __future__ import annotations

import math
from dataclasses import dataclass
from typing import Optional, Tuple, Union, List

import torch
from torch import nn
import torch.nn.functional as F


class RotaryEmbedding(nn.Module):
    def __init__(self, head_dim: int, min_timescale: int, max_timescale: int):
        super().__init__()
        if head_dim % 2 != 0:
            raise ValueError("RoPE dimension must be even")
        half_dim = head_dim // 2
        fraction = (2.0 * torch.arange(0, half_dim)) / head_dim
        timescale = min_timescale * (max_timescale / min_timescale) ** fraction
        inv_freq = 1.0 / timescale
        self.register_buffer("inv_freq", inv_freq.to(torch.float32), persistent=False)

    def forward(self, x: torch.Tensor, position_ids: torch.Tensor) -> torch.Tensor:
        pos = position_ids.to(self.inv_freq.dtype)
        freqs = torch.einsum("...i,j->...ij", pos, self.inv_freq)
        emb = torch.cat((freqs, freqs), dim=-1)
        while emb.dim() < x.dim():
            emb = emb.unsqueeze(-2)
        cos = emb.cos().to(x.dtype)
        sin = emb.sin().to(x.dtype)
        x1, x2 = torch.chunk(x, 2, dim=-1)
        rotated = torch.cat((-x2, x1), dim=-1)
        return (x * cos) + (rotated * sin)


def _rotate_half(x: torch.Tensor) -> torch.Tensor:
    x1 = x[..., ::2]
    x2 = x[..., 1::2]
    return torch.stack((-x2, x1), dim=-1).reshape_as(x)


def _get_activation(name: str) -> nn.Module:
    name = name.lower()
    if name in ("silu", "swish", "swiglu"):
        return nn.SiLU()
    if name in ("gelu", "geglu"):
        return nn.GELU()
    if name == "relu":
        return nn.ReLU()
    if name == "linear":
        return nn.Identity()
    raise ValueError(f"Unsupported activation {name}")


@dataclass
class AttentionShape:
    dim: int
    heads: int
    kv_heads: int
    head_dim: int
    rope_min: int
    rope_max: int
    apply_rope: bool


class Attention(nn.Module):
    """Byte-for-byte port of dia_v2 Attention.forward_incremental."""

    def __init__(self, config: DiaConfig, dim: int, compute_dtype: torch.dtype) -> None:
        super().__init__()
        dec = config.model.decoder
        self.num_query_heads = dec.gqa_query_heads
        self.num_kv_heads = dec.kv_heads
        self.head_dim = dec.gqa_head_dim
        self.num_gqa_groups = self.num_query_heads // max(self.num_kv_heads, 1)
        self.compute_dtype = compute_dtype
        self.q_proj = nn.Linear(dim, self.num_query_heads * self.head_dim, bias=False)
        self.k_proj = nn.Linear(dim, self.num_kv_heads * self.head_dim, bias=False)
        self.v_proj = nn.Linear(dim, self.num_kv_heads * self.head_dim, bias=False)
        self.o_proj = nn.Linear(self.num_query_heads * self.head_dim, dim, bias=False)
        eps = config.model.normalization_layer_epsilon
        self.q_norm = nn.RMSNorm(self.head_dim, eps=eps, dtype=torch.float32)
        self.k_norm = nn.RMSNorm(self.head_dim, eps=eps, dtype=torch.float32)
        self.rotary = RotaryEmbedding(
            self.head_dim,
            config.model.rope_min_timescale,
            config.model.rope_max_timescale,
        )

    def forward_incremental(
        self,
        x: torch.Tensor,
        pos: Optional[torch.Tensor],
        cache_slot,
    ) -> Tuple[torch.Tensor, object]:
        B, T, _ = x.shape
        if T != 1:
            raise ValueError("Attention expects sequence length 1 during decoding")
        orig_dtype = x.dtype
        q_proj = self._project_heads(self.q_proj, x, self.num_query_heads)
        k_proj = self._project_heads(self.k_proj, x, self.num_kv_heads)
        v_proj = self._project_heads(self.v_proj, x, self.num_kv_heads)
        q_proj = self.q_norm(q_proj)
        k_proj = self.k_norm(k_proj)
        if pos is not None:
            q_proj = self.rotary(q_proj, pos)
            k_proj = self.rotary(k_proj, pos)
        q = q_proj.transpose(1, 2)
        k = k_proj.transpose(1, 2)
        v = v_proj.transpose(1, 2)
        if cache_slot is not None:
            k_cache, v_cache, attn_mask = cache_slot.write_and_view(k, v)
        else:
            k_cache, v_cache = k, v
            attn_mask = None
        attn = F.scaled_dot_product_attention(
            q,
            k_cache,
            v_cache,
            scale=1.0,
            attn_mask=attn_mask,
            enable_gqa=self.num_gqa_groups > 1,
        )
        attn = attn.transpose(1, 2).contiguous()
        flat = attn.reshape(B, T, self.num_query_heads * self.head_dim)
        out = self.o_proj(flat.to(torch.float32))
        return out.to(orig_dtype), cache_slot

    def _project_heads(self, layer: nn.Linear, x: torch.Tensor, heads: int) -> torch.Tensor:
        proj = layer(x.to(torch.float32))
        B, T, _ = proj.shape
        proj = proj.view(B, T, heads, self.head_dim)
        return proj.to(self.compute_dtype)

    def forward(
        self,
        x: torch.Tensor,
        positions: Optional[torch.Tensor],
        cache=None,
    ) -> Tuple[torch.Tensor, Tuple[torch.Tensor, torch.Tensor]]:
        return self.forward_incremental(x, positions, cache)



class MultiStreamEmbedding(nn.Module):
    """Port of dia_v2 MultiStreamEmbed."""

    def __init__(
        self,
        vocab_size: int,
        dim: int,
        pad_id: int,
        *,
        output_dtype: torch.dtype,
        low_rank_dim: Optional[int] = None,
    ) -> None:
        super().__init__()
        self.pad_id = pad_id
        self.dtype = output_dtype
        base_dim = low_rank_dim if low_rank_dim is not None else dim
        self.embedding = nn.Embedding(vocab_size, base_dim)
        self.main_proj = nn.Linear(base_dim, dim, bias=False)
        self.second_proj = nn.Linear(base_dim, dim, bias=False)

    def forward(self, main_inputs: torch.Tensor, second_inputs: torch.Tensor) -> torch.Tensor:
        main_inputs = main_inputs.long()
        second_inputs = second_inputs.long()
        if self.pad_id is not None:
            second_is_pad = second_inputs == self.pad_id
        else:
            second_is_pad = torch.zeros_like(second_inputs, dtype=torch.bool)
        use_second = ~second_is_pad
        emb_main = self.embedding(main_inputs)
        emb_second = self.embedding(second_inputs)
        out_main = self.main_proj(emb_main.to(torch.float32))
        out_second = self.second_proj(emb_second.to(torch.float32))
        zeros = torch.zeros_like(out_second)
        y = out_main + torch.where(use_second.unsqueeze(-1), out_second, zeros)
        target_dtype = self.dtype if self.dtype is not None else y.dtype
        return y.to(target_dtype)


class Mlp(nn.Module):
    """Port of dia_v2 MlpBlock (two-activation gated MLP)."""

    def __init__(
        self,
        dim: int,
        hidden: int,
        compute_dtype: torch.dtype,
        activations: Sequence[str],
    ) -> None:
        super().__init__()
        if len(activations) != 2:
            raise ValueError("Mlp expects two activation functions.")
        self.dtype = compute_dtype
        self.hidden = hidden
        self.branch_count = len(activations)
        self.wi = nn.Linear(dim, self.branch_count * hidden, bias=False)
        self.wo = nn.Linear(hidden, dim, bias=False)
        self.activation_fns = [_get_activation(activations[0]), _get_activation(activations[1])]

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        proj = self.wi(x.to(torch.float32))
        proj = proj.view(*x.shape[:-1], self.branch_count, self.hidden).to(self.dtype)
        gate, up = proj.unbind(dim=-2)
        hidden = self.activation_fns[0](gate) * self.activation_fns[1](up)
        out = self.wo(hidden.to(torch.float32))
        return out.to(self.dtype)