Skip to content

vllm_gaudi.models.deepseek_v2

_get_hpu_llama_4_scaling 

_get_hpu_llama_4_scaling(
    original_max_position_embeddings: int,
    scaling_beta: float,
    positions: Tensor,
) -> Tensor
Source code in vllm_gaudi/models/deepseek_v2.py 
 5
 6
 7
 8
 9
10
11
12
def _get_hpu_llama_4_scaling(original_max_position_embeddings: int, scaling_beta: float,
                             positions: torch.Tensor) -> torch.Tensor:
    scaling = 1 + scaling_beta * torch.log(1 + torch.floor(positions / original_max_position_embeddings))
    # Broadcast over num_heads and head_dim
    scaling = scaling[..., None, None]

    # Squeeze dimension of scaling factor to match expected shape on HPU
    return scaling.reshape(-1, *scaling.shape[-2:])