BBQ Support in SGLang

[akhauriyash]

Summary

Adds native SGLang serving support for XllmForCausalLM, used by the 375B BBQ/xLLM MoE checkpoint.

Changes

• Add sglang.srt.models.xllm.
• Register EntryClass = XllmForCausalLM.
• Use SGLang native TopK.forward_native for MoE routing.
• Preserve xLLM partial-RoPE behavior.
• Keep the 375B MoE path single-stream while still allowing CUDA graph capture.
• Increase the unbalanced model-loading timeout for large multi-node loads.
• Add unit tests for registration, expert-location metadata, partial RoPE, and native TopK routing.

Why

This avoids relying on runtime file overlays and fixes the subtle long-decode failure mode seen with the generic TopK path on 375B MoE serving.

Validation

• python3 -m py_compile python/sglang/srt/models/xllm.py python/sglang/test/test_xllm_model.py
• git diff --check

[nightlessbaron]

LGTM!

[nightlessbaron]

@akhauriyash, can you please add tests to ensure that each of the sub-modules work?

need tests

[akhauriyash]

@akhauriyash, can you please add tests to ensure that each of the sub-modules work?

Which sub-module? SGLang? I think you mean submodules in RL360 right?

I havent landed my changes in RL360 yet