[Platform] import activation_quant_fusion for CUDA only

[wangxiyuan]

Purpose

The commit #23671 add the custom init logic like SILU_MUL_OP = torch.ops._C.silu_and_mul.default. But it doesn't work on the platform which doesn't support custom ops. It will raise error when setup vLLM.

AttributeError: '_OpNamespace' '_C' object has no attribute 'silu_and_mul'

This PR let activation_quant_fusion be imported only for CUDA case

Test Plan

This is a change for custom platform. No need for new test. and all test should not break. vLLM Ascend e2e test will be happy after this change.

Test Result

Custom platform works well then.

[gemini-code-assist]

Code Review

The pull request aims to fix an AttributeError on platforms without custom ops by adding hasattr checks. While the intention is correct, the implementation is incomplete and introduces new runtime errors because other parts of the code unconditionally assume these ops are available. The pattern registration logic needs to be made conditional to fully resolve the issue.

[elvischenv]

https://github.com/vllm-project/vllm/blob/main/vllm/compilation/fusion.py#L41-L54
Will these lines raise error for you?

[wangxiyuan]

@elvischenv there is a platform check when import the file, so it's fine with custom platform

vllm/vllm/compilation/pass_manager.py

Lines 10 to 12 in 006477e

	if current_platform.is_cuda_alike():
	from .fusion import FusionPass
	from .fusion_attn import AttnFusionPass

And I'm not sure if activation_quant_fusion is only used for CUDA, if yes, it can be fixed by this way as well.

[jikunshang]

Thanks for fixing this! maybe we need some refactor to avoid op register issue on multiple platforms.

[ProExpertProg]

This is very ugly, but I understand that we can't set these globals due to other platforms. I think we've already dealt with this in other files and in a nicer way, could you look around briefly?

[ProExpertProg]

Yeah I see the link now, can you just make the pass a conditional import like it's done for fusion?

[elvischenv]

And I'm not sure if activation_quant_fusion is only used for CUDA, if yes, it can be fixed by this way as well.

Yes, this is only for cuda. Let's follow this conditional import as @ProExpertProg mentioned. Thanks.

[wangxiyuan]

@elvischenv @ProExpertProg Done, and tested with vllm-ascend again, it works