Fix aclnnRepeatInterleaveIntWithDim error on NPU for get_1d_rotary_pos_embed

[ZhengKai91]

What does this PR do?

While running diffusers on NPU, the following error was encountered:
RuntimeError: call aclnnRepeatInterleaveIntWithDim failed, detail:EZ1001: 2025-02-18-15:02:05.865.233 self not implemented for DT_DOUBLE, should be in dtype support list [DT_UINT8,DT_INT8,DT_INT16,DT_INT32,DT_INT64,DT_BOOL,DT_FLOAT16,DT_FLOAT,DT_BFLOAT16,].
This error occurs specifically on NPU due to the unsupported data type DT_DOUBLE for the repeat_interleave operation. The error is triggered in the get_1d_rotary_pos_embed function in diffusers/models/embeddings.py
specifically in the following line:
freqs_cos = freqs.cos().repeat_interleave(2, dim=1).float() # [S, D]

Solution
To resolve this issue, I explicitly cast freqs to float before the cos operation. This ensures that the data type is supported by the repeat_interleave operation on NPU. The modified code is as follows:
freqs = torch.outer(pos, freqs).float() # type: ignore # [S, D/2]
if use_real and repeat_interleave_real:
# flux, hunyuan-dit, cogvideox
freqs_cos = freqs.cos().repeat_interleave(2, dim=1).float() # [S, D]
This change ensures compatibility with the data types supported by NPU

Testing
I have tested the modified code on NPU and confirmed that this change resolves the error without affecting the normal operation of other functionalities.
Documentation Updates
No documentation updates were made, as this is a bug fix specific to NPU and does not introduce new features or significant changes.
Dependencies
No new dependencies are introduced by this change.
Issue Fixed
Fixes the aclnnRepeatInterleaveIntWithDim error on NPU for the get_1d_rotary_pos_embed function.

Pre-Submission Checklist
[x] This PR fixes a bug specific to NPU.
[x] I have read the contributor guideline.
[x] I have read the philosophy doc.
[x] No discussion or approval via GitHub issue or forum was needed for this change.
[x] No documentation updates were necessary.
[x] No new tests were added, as this is a bug fix.
Reviewers
@yiyixuxu @sayakpaul

[hlky]

Thanks @ZhengKai91. We are aiming to improve alternative device type support and appreciate any contributions towards that goal!

[hlky]

Suggested change

	freqs = torch.outer(pos, freqs).float() # type: ignore # [S, D/2]
	freqs = torch.outer(pos, freqs) # type: ignore # [S, D/2]
	is_npu = freqs.device.type == "npu"
	if is_npu:
	freqs = freqs.float()

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.