optimum/habana/transformers/models/deepseek_v2/modeling_deepseek_v2.py [358:408]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    def __init__(
        self,
        dim,
        max_position_embeddings=2048,
        base=10000,
        device=None,
        scaling_factor=1.0,
        original_max_position_embeddings=4096,
        beta_fast=32,
        beta_slow=1,
        mscale=1,
        mscale_all_dim=0,
    ):
        self.scaling_factor = scaling_factor
        self.original_max_position_embeddings = original_max_position_embeddings
        self.beta_fast = beta_fast
        self.beta_slow = beta_slow
        self.mscale = mscale
        self.mscale_all_dim = mscale_all_dim
        super().__init__(dim, max_position_embeddings, base, device)

    def _set_cos_sin_cache(self, seq_len, device, dtype):
        self.max_seq_len_cached = seq_len
        dim = self.dim

        freq_extra = 1.0 / (self.base ** (torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim))
        freq_inter = 1.0 / (
            self.scaling_factor * self.base ** (torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim)
        )

        low, high = yarn_find_correction_range(
            self.beta_fast,
            self.beta_slow,
            dim,
            self.base,
            self.original_max_position_embeddings,
        )
        inv_freq_mask = 1.0 - yarn_linear_ramp_mask(low, high, dim // 2).to(device=device, dtype=torch.float32)
        inv_freq = freq_inter * (1 - inv_freq_mask) + freq_extra * inv_freq_mask
        self.register_buffer("inv_freq", inv_freq, persistent=False)

        t = torch.arange(seq_len, device=device, dtype=torch.float32)

        freqs = torch.outer(t, inv_freq)

        _mscale = float(
            yarn_get_mscale(self.scaling_factor, self.mscale)
            / yarn_get_mscale(self.scaling_factor, self.mscale_all_dim)
        )

        emb = torch.cat((freqs, freqs), dim=-1)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -



optimum/habana/transformers/models/deepseek_v3/modeling_deepseek_v3.py [274:324]:
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -
    def __init__(
        self,
        dim,
        max_position_embeddings=2048,
        base=10000,
        device=None,
        scaling_factor=1.0,
        original_max_position_embeddings=4096,
        beta_fast=32,
        beta_slow=1,
        mscale=1,
        mscale_all_dim=0,
    ):
        self.scaling_factor = scaling_factor
        self.original_max_position_embeddings = original_max_position_embeddings
        self.beta_fast = beta_fast
        self.beta_slow = beta_slow
        self.mscale = mscale
        self.mscale_all_dim = mscale_all_dim
        super().__init__(dim, max_position_embeddings, base, device)

    def _set_cos_sin_cache(self, seq_len, device, dtype):
        self.max_seq_len_cached = seq_len
        dim = self.dim

        freq_extra = 1.0 / (self.base ** (torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim))
        freq_inter = 1.0 / (
            self.scaling_factor * self.base ** (torch.arange(0, dim, 2, dtype=torch.float32, device=device) / dim)
        )

        low, high = yarn_find_correction_range(
            self.beta_fast,
            self.beta_slow,
            dim,
            self.base,
            self.original_max_position_embeddings,
        )
        inv_freq_mask = 1.0 - yarn_linear_ramp_mask(low, high, dim // 2).to(device=device, dtype=torch.float32)
        inv_freq = freq_inter * (1 - inv_freq_mask) + freq_extra * inv_freq_mask
        self.register_buffer("inv_freq", inv_freq, persistent=False)

        t = torch.arange(seq_len, device=device, dtype=torch.float32)

        freqs = torch.outer(t, inv_freq)

        _mscale = float(
            yarn_get_mscale(self.scaling_factor, self.mscale)
            / yarn_get_mscale(self.scaling_factor, self.mscale_all_dim)
        )

        emb = torch.cat((freqs, freqs), dim=-1)
- - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -