src/model/llama_condense_monkey_patch.py

# Code adapted from https://huggingface.co/kaiokendev/superhot-13b-8k-no-rlhf-test/blob/main/llama_rope_scaled_monkey_patch.py

from functools import partial

import torch
import transformers
import transformers.models.llama.modeling_llama


class CondenseRotaryEmbedding(torch.nn.Module):
    def __init__(
        self, dim, ratio, max_position_embeddings=2048, base=10000, device=None
    ):
        super().__init__()
        inv_freq = 1.0 / (base ** (torch.arange(0, dim, 2).float().to(device) / dim))
        self.register_buffer("inv_freq", inv_freq)

        # Build here to make `torch.jit.trace` work.
        self.ratio = ratio
        max_position_embeddings *= ratio
        self.max_seq_len_cached = max_position_embeddings
        # print(f"Monkey Patching condense ratio {ratio}")
        t = (
            torch.arange(
                self.max_seq_len_cached,
                device=self.inv_freq.device,
                dtype=self.inv_freq.dtype,
            )
            / 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
        )

    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.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),
        )


def replace_llama_with_condense(ratio):
    transformers.models.llama.modeling_llama.LlamaRotaryEmbedding = partial(
        CondenseRotaryEmbedding, ratio=ratio
    )