[NSA] Fall back to fast_hadamard_transform when sgl_kernel lacks the symbol

[Fridge003]

Summary

• On non-HIP / non-SM103 paths, nsa_indexer.rotate_activation does from sgl_kernel import hadamard_transform. Older sgl_kernel builds (e.g. 0.3.21) don't export this symbol, which causes ImportError at forward time and crashes the scheduler. Wrap the import in try/except and fall back to fast_hadamard_transform (already used on HIP / SM103), so older / partial sgl_kernel builds keep working.

Why

• Reproducible on the official lmsysorg/sglang:deepseek-v4-grace-blackwell image only when running on non-GB300 CUDA hardware where _is_sm103=False. On GB300 it's masked because _is_sm103=True already routes to fast_hadamard_transform. But any non-Blackwell / non-GB300 CUDA platform hits this:

File ".../sglang/srt/layers/attention/nsa/nsa_indexer.py", line ..., in rotate_activation
    from sgl_kernel import hadamard_transform
ImportError: cannot import name 'hadamard_transform' from 'sgl_kernel'
[..] Received sigquit from a child process. It usually means the child failed.

• fast_hadamard_transform is already an installed dependency on those code paths, so the fallback adds no new requirement.

Diff

def rotate_activation(x: torch.Tensor) -> torch.Tensor:
    if _is_hip or _is_sm103:
        from fast_hadamard_transform import hadamard_transform
    else:
        try:
            from sgl_kernel import hadamard_transform
        except ImportError:
            from fast_hadamard_transform import hadamard_transform
    ...

Test plan

• On a GB300 pod with lmsysorg/sglang:deepseek-v4-grace-blackwell (sgl-kernel 0.3.21 — no hadamard_transform): _is_sm103=True path still chosen; gsm8k 20-shot sanity acc 0.949 (Flash low-latency) — unchanged.
• On the same pod when forcing the non-SM103 branch (older internal image fridge003/sglang:final-gb300 where _is_sm103=False because the source predates SM103 detection): without this patch, the scheduler crashed at first forward; with the equivalent patch applied locally, gsm8k 20-shot acc 0.950 — fix confirmed.
• CI on deepseek_v4 (let upstream run it).

🤖 Generated with Claude Code

[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!