Skip to content

fix(dlinfer): remove CUDA hardcode in NTK rotary embedding for non-CUDA accelerators #4986

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
lvhan028 merged 2 commits into from
Sep 22, 2026
+3 −1
Merged

fix(dlinfer): remove CUDA hardcode in NTK rotary embedding for non-CUDA accelerators #4986

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
81a6e19
fix(dlinfer): remove CUDA hardcode in NTK rotary embedding
li-lizhe Sep 18, 2026
887587e
fix(dlinfer): cache pos_freq_scaling device to avoid per-forward copy
li-lizhe Sep 18, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 3 additions & 1 deletion lmdeploy/pytorch/backends/dlinfer/rotary_embedding.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -76,14 +76,16 @@ class DlinferLlamaDynamicNTKScalingRotaryEmbedding(LlamaDynamicNTKScalingRotaryE
def __init__(self, dim: int, base: int = 10000, scaling_factor: float = 1.0, max_position_embeddings: int = 2048):
super().__init__(dim, base, scaling_factor, max_position_embeddings)
self.dim_scale_ratio = self.dim / (self.dim - 2)
self.pos_freq_scaling = torch.arange(0, self.dim, 2, dtype=torch.int64).float().cuda() / self.dim
self.pos_freq_scaling = torch.arange(0, self.dim, 2, dtype=torch.int64).float() / self.dim
self.scale_offset = self.scaling_factor - 1
self.pos_scale_factor = self.scaling_factor / \
self.max_position_embeddings

def _ntk_inv_freq(self, seq_len: torch.Tensor):
"""Calculate inverse frequency with NTK scaling."""
base = self.base * ((self.pos_scale_factor * seq_len) - self.scale_offset)**self.dim_scale_ratio
if self.pos_freq_scaling.device != seq_len.device:
self.pos_freq_scaling = self.pos_freq_scaling.to(seq_len.device)
inv_freq = 1.0 / (base**self.pos_freq_scaling)
return inv_freq

Expand Down
Loading