[AMD] Fix FA3 support check crash on ROCm (torch.version.cuda is None)

[bingxche]

Motivation

PR #20796 (1a8eb890f6, "Kernels community fa3", merged 2026-04-07) introduced a new unified flash attention dispatch layer (python/sglang/jit_kernel/flash_attention.py and flash_attention_v3.py). As part of this refactor, multimodal_gen's flash attention backend was changed from directly importing sgl_kernel.flash_attn.flash_attn_varlen_func to importing from the new sglang.jit_kernel.flash_attention module.

The new code path unconditionally calls _is_fa3_supported() before dispatching, which checks:

return (torch.version.cuda >= "12.3") and (...)

On ROCm (AMD GPUs), torch.version.cuda is None, causing:

TypeError: '>=' not supported between instances of 'NoneType' and 'str'

This crashes all diffusion tests that use Qwen-Image or Z-Image-Turbo models during the TextEncodingStage, because the text encoder's attention layer goes through the new FA3 code path.

Impact

All 3 multimodal-gen test jobs in the AMD AITER Scout #29 workflow failed:

Job	Failed Test	Model
multimodal-gen-test-1-gpu-amd (part 0)	qwen_image_t2i	Qwen/Qwen-Image
multimodal-gen-test-1-gpu-amd (part 1)	qwen_image_t2i_cache_dit_enabled	Qwen/Qwen-Image
multimodal-gen-test-2-gpu-amd (part 0)	fsdp-inference	Tongyi-MAI/Z-Image-Turbo

Each job retried 6 times and exhausted all retries with the same TypeError.

Fix

Guard _is_fa3_supported() to return False when torch.version.cuda is None (i.e., on ROCm), since FA3 is a CUDA-only feature.

Test Plan

• Verify multimodal-gen-test-1-gpu-amd passes (Qwen-Image T2I tests)
• Verify multimodal-gen-test-2-gpu-amd passes (Z-Image-Turbo FSDP test)
• No impact on CUDA paths (the guard only triggers when torch.version.cuda is None)

Made with Cursor

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[bingxche]

test https://github.com/sgl-project/sglang/actions/runs/24145031904

[polisettyvarma]

@bingxche it's a problem for XPU also
when can this PR be marked ready for review ?

[bingxche]

@bingxche it's a problem for XPU also when can this PR be marked ready for review ?

Waiting for CI test https://github.com/sgl-project/sglang/actions/runs/24124568923/job/70386014759, after the test passed I will mark it as ready for review.

[polisettyvarma]

@bingxche it's a problem for XPU also when can this PR be marked ready for review ?

Waiting for CI test https://github.com/sgl-project/sglang/actions/runs/24124568923/job/70386014759, after the test passed I will mark it as ready for review.

@bingxche seems test failed, please check