[desmond-intel]

Motivation

Baichuan2-13B / Baichuan-13B use ALiBi and skip RoPE. baichuan.py passes the ALiBi slopes to attention only on NPU, so on XPU the model runs with no position information and degenerates (0 0 0 ...). Baichuan2-13B-Chat gsm8k is 0.00 on B60 (vLLM RTX 6000 Blackwell: 53.5).

Modifications

• models/baichuan.py: pass slopes on XPU too, as fp32.
• layers/attention/xpu_backend.py: forward_extend / forward_decode accept slopes and forward it as alibi_slopes to flash_attn_with_kvcache, next to the existing sinks kwarg.

Other models and backends are unchanged. On XPU with --attention-backend triton, Baichuan-13B now raises at the first forward (triton takes no slopes) instead of running without position information.

Depends on sgl-project/sgl-kernel-xpu#573 (alibi_slopes in the FMHA kernels) and the sgl-kernel-xpu pin bump that follows it.

Accuracy Tests

Baichuan2-13B-Chat, tp2, --attention-backend intel_xpu, gsm8k 5-shot, B60:

	docs	flexible	strict
main	50	0.00	0.00
this PR	50	0.52	0.52
this PR	1319	0.5474	0.5459

Reference: vLLM RTX 6000 Blackwell 53.5, paper 56.8 (4-shot).

python -m sglang.launch_server --model-path baichuan-inc/Baichuan2-13B-Chat --trust-remote-code \
  --attention-backend intel_xpu --tp-size 2

Speed Tests and Profiling

No speed impact on other models; the slopes are only passed for ALiBi Baichuan on XPU.

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests. Not added here: the kernel PR carries test_alibi_attention, and the model-level change is covered by the accuracy run above.
• Update documentation according to Write documentations. Not needed.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #39912544389
Latest PR Test (Extra): ❌ Run #39912544313
Latest PR Test (AMD ROCm 10): ❌ Run #39912544319