[em-dreyer]

Motivation

The nightly test_ling_2_6_flash.py (8-GPU H200 and B200) has failed GSM8K since #44904. On 10/21 (8e3e2980d) it scored 0.39 on H200 and 0.35 on B200, against a 0.825 threshold. The 10/17 nightly scored 0.87.

• H200: https://github.com/sgl-project/sglang/actions/runs/39614482141/job/119271892494
• B200: https://github.com/sgl-project/sglang/actions/runs/39614482141/job/119271892388

Root cause:

1. The BailingMoeV2_5ForCausalLM checkpoints (Ling-2.6-flash, Ling-2.5-1T, Ling-2.6-1T, Ring-2.5-1T) have no norm_topk_prob key. bailing_moe_linear.py reads getattr(config, "norm_topk_prob", False), so the router gets renormalize=False.
2. The reference modeling_bailing_moe_v2_5.py always renormalizes the top-k weights, then multiplies by routed_scaling_factor (2.5):

   topk_weight = scores / (scores.sum(dim=-1, keepdim=True) + 1e-20) if self.top_k > 1 else scores
   topk_weight = topk_weight * self.routed_scaling_factor

3. Ling-2.6-flash (256 experts, top-8) used to take the FlashInfer fused_topk_deepseek path, which always renormalizes regardless of the flag. That hid the wrong default.
4. [Fix] moe: Renormalize per-rank shared-slot routing over routed experts only #44904 correctly gated that path on renormalize, so Ling moved to a path that respects renormalize=False. The top-k weights no longer sum to 1, the MoE output is mis-scaled at every layer, and GSM8K collapsed.

#44904 is correct and should stay; the bug is the model's default.

Modifications

bailing_moe_linear.py: norm_topk_prob now defaults to True, matching the reference modeling. Configs that set the key explicitly are unaffected.

Accuracy Tests

The CI test file was run unchanged (test_ling_2_6_flash.py: TP=4, NEXTN MTP, extra_buffer, 200-question 5-shot GSM8K) on 4×B200:

Commit	GSM8K	Mean accept len
4eb606b2ae (10/17 nightly, last green)	0.87	3.40
3a059de94c (parent of #44904)	0.86	—
48ea486928 (#44904)	0.355	2.38
8e3e2980db (10/21 nightly)	0.43	2.68
8e3e2980db + this change	0.87	3.40

The first bad commit was found by bisecting 62cee1b15c..8e3e2980db. The fixed run applied this same one-line change on top of 8e3e2980db; this branch is based on a newer main.

Not covered here: bailing_moe_v3.py has the same getattr(config, "norm_topk_prob", False) default. I haven't checked whether its reference modeling also always renormalizes.

Checklist

• Format your code according to the Format code with pre-commit.
• Provide accuracy and speed benchmark results according to Test the accuracy.

cc @ch-wan (#44904)

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ❌ Run #39671654604
Latest PR Test (Extra): ❌ Run #39671654736
Latest PR Test (AMD ROCm 10): ❌ Run #39671655084

[ch-wan]

LGTM