[Bugfix] Add forward_xpu to XDRotaryEmbedding for HunyuanOCR on XPU (#52174)

Signed-off-by: Jakub Byczkowski <jakub.byczkowski@intel.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
This commit is contained in:
Jakub Byczkowski
2026-08-18 07:54:05 +08:00
committed by GitHub
co-authored by Copilot
parent 49fb2ee348
commit c296cf8259
@@ -137,6 +137,18 @@ class XDRotaryEmbedding(DynamicNTKAlphaRotaryEmbedding):
key = torch.cat((key_rot, key_pass), dim=-1).reshape(key_shape)
return query, key
def forward_xpu(
self,
positions: torch.Tensor,
query: torch.Tensor,
key: torch.Tensor | None = None,
offsets: torch.Tensor | None = None,
) -> tuple[torch.Tensor, torch.Tensor | None]:
# No fused XPU kernel for XD-sectioned RoPE; the base
# RotaryEmbedding.forward_xpu calls the generic C++ kernel, which
# rejects the 2-D [4, num_tokens] xdrope positions. Use the native path.
return self.forward_native(positions, query, key, offsets)
@staticmethod
def get_next_input_positions(
context_len: int,