from functools import partial
import torch
import transformers
import transformers.models.llama.modeling_llama
[docs]
class CondenseRotaryEmbedding(torch.nn.Module):
def __init__(self, dim, pi_ratio, ntk_ratio, max_position_embeddings=2048, base=10000, device=None):
super().__init__()
[docs]
self.ntk_ratio = ntk_ratio
max_position_embeddings *= ntk_ratio
base = base * ntk_ratio ** (dim / (dim-2)) #Base change formula
inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
self.register_buffer("inv_freq", inv_freq)
[docs]
self.pi_ratio = pi_ratio
max_position_embeddings *= pi_ratio
[docs]
self.max_seq_len_cached = max_position_embeddings
t = torch.arange(self.max_seq_len_cached, device=self.inv_freq.device, dtype=self.inv_freq.dtype) / pi_ratio
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1)
dtype = torch.get_default_dtype()
self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(dtype), persistent=False)
[docs]
def forward(self, x, seq_len=None):
# x: [bs, num_attention_heads, seq_len, head_size]
# This `if` block is unlikely to be run after we build sin/cos in `__init__`. Keep the logic here just in case.
if seq_len > self.max_seq_len_cached:
self.max_seq_len_cached = seq_len
t = torch.arange(self.max_seq_len_cached, device=x.device, dtype=self.inv_freq.dtype) / self.pi_ratio
freqs = torch.einsum("i,j->ij", t, self.inv_freq)
# Different from paper, but it uses a different permutation in order to obtain the same calculation
emb = torch.cat((freqs, freqs), dim=-1).to(x.device)
self.register_buffer("cos_cached", emb.cos()[None, None, :, :].to(x.dtype), persistent=False)
self.register_buffer("sin_cached", emb.sin()[None, None, :, :].to(x.dtype), persistent=False)
return (
self.cos_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
self.sin_cached[:, :, :seq_len, ...].to(dtype=x.dtype),
)
[docs]
def replace_llama_with_condense(pi_ratio, ntk_ratio):
transformers.models.llama.modeling_llama.LlamaRotaryEmbedding = partial(CondenseRotaryEmbedding, pi_ratio=pi_ratio, ntk_ratio=ntk_ratio)
