[XPU][Diffusion] Enable MiniMax H3 on XPU platforms

[jianan-gu]

This PR adds key changes to cover the enabling for Minimax H3 model specific on XPU platforms, including

• Mapping to XPU attention backend with optimized kernel from sgl-kernel-xpu
• Necessary device dispatching/autocast codes
• Runtime XPU memory D2H/H2D fix

Tested serving pipelines include:

• TP + layer_wise offload for both encoder/dit/vae ( similar as cookbook on cuda 5090: https://docs.sglang.io/cookbook/diffusion/MiniMax/MiniMax-H3#3-serve-minimax-h3, but preferring no offload (only encoder) when device memory is enough, i.e., x4 or x8)

Next plans (as follow-up PRs, for further optimizations).

• Apply more fusions (like QKnorm_RoPE and silu_and_mul, to be on par with cuda fusion pathes)
• More serving pipelines/parallelism to be explored and evaluated.
• XPU FA backend with non-contiguous support

Tested example:

• T2VA example from cookbook: https://docs.sglang.io/cookbook/diffusion/MiniMax/MiniMax-H3#4-generate-video-and-audio

Result:

cookbook.mp4

---

CI States

Latest PR Test (Base): ✅ Run #34183716549
Latest PR Test (Extra): ✅ Run #34183716399
Latest PR Test (AMD ROCm 7.2): ❌ Run #34183716525

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.

[jianan-gu]

/tag-and-rerun-ci

[jianan-gu]

/tag-run-ci-label extra

[jianan-gu]

/rerun-failed-ci

[jianan-gu]

/rerun-failed-ci

[jianan-gu]

/rerun-failed-ci

[jianan-gu]

/rerun-failed-ci

[jianan-gu]

/rerun-failed-ci

[jianan-gu]

Hi @mickqian could you help revisit and take a look on this PR? Thanks.
The PR changes much simpler now as we have rebased with main that like NPUs are doing similar refactoring for H3.
The CI is passed for CUDA, and the AMD CI test failures are not related to this PR, also happens on main: https://github.com/sgl-project/sglang/actions/runs/33456144533

[jianan-gu]

Hi @mingfeima, after rebase from main, this PR is XPU specific now, all previous common components changes now we are leveraging the common path from main, by just adding is_xpu. And the main CI (CUDA) is passed, with some failures (AMD suite) also happen on main. Shall we proceed? Thanks

[mingfeima]

@jianan-gu rebase and fix the conflict.

[jianan-gu]

/rerun-failed-ci