| 80 | |
| 81 | |
| 82 | class RotaryEmbedding(torch.nn.Module): |
| 83 | # Extracted from: https://github.com/EleutherAI/gpt-neox |
| 84 | def __init__(self, dim, config, base=10000,precision=torch.half): |
| 85 | super().__init__() |
| 86 | self.config = config |
| 87 | self.dim = dim |
| 88 | self.base = base |
| 89 | self.max_seq_len_cached = None |
| 90 | self.cos_cached = None |
| 91 | self.sin_cached = None |
| 92 | self.precision = precision |
| 93 | |
| 94 | def get_mscale(self, scale=1): |
| 95 | if scale <= 1: |
| 96 | return 1.0 |
| 97 | return 0.1 * math.log(scale) + 1.0 |
| 98 | |
| 99 | def get_ntk_alpha(self, true_seq_len): |
| 100 | context_value = math.log(true_seq_len / 4096, 2) + 1 |
| 101 | # ntk_alpha = 2 ** context_value - 1 |
| 102 | ntk_alpha = 2 ** math.ceil(context_value) - 1 |
| 103 | ntk_alpha = max(ntk_alpha, 1) |
| 104 | return ntk_alpha |
| 105 | |
| 106 | def forward(self, x, seq_dim=0, seq_len=None): |
| 107 | if seq_len is None: |
| 108 | seq_len = x.shape[seq_dim] |
| 109 | seq_len = max(seq_len, self.config.training_seqlen) |
| 110 | ntk_alpha = self.get_ntk_alpha(seq_len) |
| 111 | mscale = float(self.get_mscale(seq_len / self.config.training_seqlen)) |
| 112 | base = self.base * ntk_alpha ** (self.dim / (self.dim - 2)) |
| 113 | inv_freq = 1.0 / (base ** (torch.arange(0, self.dim, 2, device=x.device).float( )/ self.dim )) |
| 114 | max_seq_len_cached = seq_len |
| 115 | t = torch.arange(max_seq_len_cached, device=x.device, dtype=inv_freq.dtype) |
| 116 | freqs = torch.einsum('i,j->ij', t, inv_freq) |
| 117 | # Different from paper, but it uses a different permutation in order to obtain the same calculation |
| 118 | emb = torch.cat((freqs, freqs), dim=-1).to(x.device) |
| 119 | if self.precision == torch.bfloat16: |
| 120 | emb = emb.float() |
| 121 | # [sx, 1 (b * np), hn] |
| 122 | cos_cached = mscale *emb.cos()[:, None, :].half() |
| 123 | sin_cached = mscale *emb.sin()[:, None, :].half() |
| 124 | if self.precision == torch.bfloat16: |
| 125 | cos_cached = cos_cached.bfloat16() |
| 126 | sin_cached = sin_cached.bfloat16() |
| 127 | return cos_cached[:seq_len, ...], sin_cached[:seq_len, ...] |
| 128 | |
| 129 | |
| 130 | # rotary pos emb helpers: |