perf: use fused gdn gating for qwen3.5 prefill.

[yingxudeng]

No description provided.

[gemini-code-assist]

Code Review

This pull request replaces manual PyTorch operations with a fused GDN gating kernel in the prefill path of the Qwen3 Gated Delta Net and adds a new test case for large batch sizes. Review feedback suggests optimizing tensor reshaping by using .reshape() instead of the .contiguous().view() pattern and removing redundant .squeeze(0) and .contiguous() calls on kernel outputs to improve performance in the prefill hot path.

[gemini-code-assist]

Prefer using .reshape() over the .contiguous().view() pattern. reshape() is more idiomatic in PyTorch; it returns a view if the tensor is already contiguous and only performs a copy if necessary. This avoids redundant operations and potential memory allocations if the input tensors a and b are already contiguous.

Suggested change

	gdn_params.a = a.contiguous().view({-1, a.size(-1)});
	gdn_params.b = b.contiguous().view({-1, b.size(-1)});
	gdn_params.a = a.reshape({-1, a.size(-1)});
	gdn_params.b = b.reshape({-1, b.size(-1)});

[gemini-code-assist]

The calls to .squeeze(0) and .contiguous() are redundant here. Tensors returned by custom kernels are typically contiguous, and view() can handle the reshaping directly from the kernel's output shape (whether it is [total_tokens, hidden] or [1, total_tokens, hidden]) to the target [batch, seq, hidden] shape. Removing these unnecessary calls improves performance in the prefill hot path.

Suggested change

	g = g.squeeze(0).contiguous().view({batch_size, seq_len, a.size(-1)});
	beta = beta.squeeze(0).contiguous().view({batch_size, seq_len, b.size(-1)});
	g = g.view({batch_size, seq_len, a.size(-1)});
	beta = beta.view({batch_size, seq_len, b.size(-1)});

[yingxudeng]

用融合算子前：

用融合算子后：

用融合算子后，性能略微裂化，此pr暂不合并

[yingxudeng]

修改代码使其编译出来 bs3520 对应kernel，并命中，性能有提升但是仍然不如原版。因此暂不合并。