Fix additive attention mask handling in the native NPU attention backend.

[mengchengTang]

What does this PR do?

Fixes attention mask handling in the native NPU attention backend.

PyTorch SDPA defines two mask formats:

• Boolean masks are keep masks, where True means attend.
• Floating-point masks are additive biases applied directly to attention scores.

Previously, _native_npu_attention passed every mask to _maybe_modify_attn_mask_npu, which casts the mask to boolean and inverts it for npu_fusion_attention. This works for boolean masks but loses the values and semantics of floating-point additive masks.

This PR makes the mask handling explicit:

• Floating-point masks are routed through the existing native SDPA implementation, preserving their additive values.
• Boolean masks continue using npu_fusion_attention.
• Other mask dtypes are rejected with a clear error instead of being silently converted to boolean.

No new dependencies are introduced.

Fixes #14397

Test results

Scope	Test	Before the fix	After the fix	Result
End-to-end	dg845/LTX-2.3-Diffusers with verl-omni NPU training	log_prob_diff was around 1e-3 during the first 5 training steps	log_prob_diff decreased to around 1e-5 during the first 5 training steps	✓ Passed
Minimal reproduction	Floating-point additive mask reproduction from this issue	The mask semantics were reversed and the assertion failed	The additive mask is handled correctly and the assertion passes	✓ Passed

Before submitting

• Did you use an AI agent (Claude Code, Codex, Cursor, etc.) to help with this PR? If so:
  • Did you read the Coding with AI agents guide?
  • Did you run the self-review skill on the diff?
  • Did you share the final self-review notes in the PR description or a comment?
• Did you read the contributor guideline?
• Did you read our philosophy doc? (important for complex PRs)
• Was this discussed/approved via a GitHub issue or the forum? Please add a link to it if that's the case.
• Did you make sure to update the documentation with your changes? Here are the
  documentation guidelines, and
  here are tips on formatting docstrings.
• Did you write any new necessary tests?
• Are you the author (or part of the team) of the model/pipeline (only applicable for model/pipeline related PRs)?

Who can review?

Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.

[mengchengTang]

@yiyixuxu @sayakpaul @zhtmike These two Ascend NPU fixes (#14288 and #14396) are issues I ran into while adapting verl-omni — could you take a look when you get a chance? Thanks!

[zhtmike]

It is just my personal opinion, seems it is little bit strange of this fixing, it is supposed to fix mask in npu attention instead of routing to sdpa?

And kindly ping @sayakpaul, can you take a look of this PR if you have time? Seems there is some mask handling probelm in NPU attention, and this fixing is limited in NPU scope. Thanks!

[sayakpaul]

If it's NPU specific then we should add a conditional to handle it. @zhtmike could you confirm if it works as expected on CUDA?

[zhtmike]

Hi @sayakpaul, I believe the code modification is within npu scope.

the changes only takes effective only when

transformer.set_attention_backend("_native_npu") -> routing to _native_npu_attention -> fix the mask handling in NPU.

So it should be safe for CUDA.

[zhtmike]

But a kind of strange why routing to spda instead fixing the mask directly? @mengchengTang

[mengchengTang]

@zhtmike The npu_fusion_attention called by _native_npu only supports boolean masks. Routing additive masks to SDPA prevents logic errors from forcibly converting them to boolean masks. SDPA can correctly handle additive masks.