npu qwen3.5 megatron padding_free fix

[addsubmuldiv]

Problem

For Qwen3.5 / Qwen3-Next on NPU, padding_free=False may fail in the GDN non-thd path, while padding_free=True works only because the thd path derives mask from cu_seqlens_q internally and bypasses the broken external mask chain. :contentReference[oaicite:2]{index=2}

Cause

The GDN non-thd branch reads external mask from kwargs, but on NPU the correct mask for this path is attention_mask_2d. Without preferring that field, the model may consume an incompatible mask. :contentReference[oaicite:3]{index=3}

Fix

This PR updates the Qwen GDN non-thd path to prefer attention_mask_2d on NPU and fall back to the original attention_mask behavior otherwise.

Notes

This is the model-side half of the fix. It is intended to work together with the ms-swift trainer-side change that preserves attention_mask_2d.

[gemini-code-assist]

Code Review

This pull request introduces a helper function, resolve_gdn_attention_mask, to centralize attention mask logic and add support for NPU devices. This function is then integrated into the forward passes of the Qwen3-Next and Qwen3.5 models. A critical typo was identified in qwen3_next.py where an undefined function resolve_hf_attention_mask is called instead of the new helper, which would lead to a runtime error.

[gemini-code-assist]

The function call resolve_hf_attention_mask(kwargs) appears to be a typo. The function defined earlier in this file (at line 62) is named resolve_gdn_attention_mask. Using the incorrect name will result in a NameError at runtime.

Suggested change

	attention_mask = resolve_hf_attention_mask(kwargs)
	attention_mask = resolve_gdn_attention_mask(kwargs)

[Copilot]

Pull request overview

Updates Qwen3.5 / Qwen3-Next GatedDeltaNet (GDN) non-THD attention-mask handling to avoid NPU failures when padding_free=False, by preferring attention_mask_2d on NPU and falling back to the existing attention_mask-derived behavior otherwise.

Changes:

• Add a shared resolve_gdn_attention_mask() helper that selects attention_mask_2d on NPU (fallback to legacy mask logic otherwise).
• Use resolve_gdn_attention_mask() in the Qwen3.5 MM-GPT GDN non-THD forward path.
• Update Qwen3-Next GDN non-THD forward path to use a resolver (but currently references an undefined function).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/mcore_bridge/model/mm_gpts/qwen3_5.py	Switch non-THD GDN path to use shared mask resolver.
src/mcore_bridge/model/gpts/qwen3_next.py	Introduce resolve_gdn_attention_mask() and update non-THD GDN mask resolution call site.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

resolve_hf_attention_mask is referenced here but is not defined anywhere in this module (or elsewhere in the repo). This will raise a NameError at runtime. Use the newly introduced resolve_gdn_attention_mask(kwargs) here (or add/rename the helper consistently if a different resolver is intended).

Suggested change

	attention_mask = resolve_hf_attention_mask(kwargs)
	attention_mask = resolve_gdn_attention_mask(kwargs)

[Jintao-Huang]

thanks! lint 过一下