Fix Kimi K2.6 compressed-tensors MoE w2 scale loading

[bingxche]

Summary

• Treat missing/false/static actorder as non-actorder in compressed-tensors WNA16 MoE.
• Avoid allocating/loading full TP-width w2_weight_scale for checkpoints whose w2 scales are already per-rank, such as Kimi-K2.6.
• Keep full w2 scale loading for real actorder/g_idx cases.

Validation

• PYTHONPYCACHEPREFIX=/tmp/sglang-pycache python3 -m py_compile python/sglang/srt/layers/quantization/compressed_tensors/schemes/compressed_tensors_wNa16_moe.py
• pre-commit hooks from git commit

CI

• MI35x Kimi-K2.6: https://github.com/sgl-project/sglang/actions/runs/27938672153

---

CI States

Latest PR Test (Base): ❌ Run #28089355031
Latest PR Test (Extra): ❌ Run #28089354818

[gemini-code-assist]

Code Review

This pull request updates compressed_tensors_wNa16_moe.py to properly handle self.actorder when it is an Enum or a mixed-case string. The reviewer suggests normalizing self.actorder in-place on the instance rather than using a local variable, ensuring that other references to self.actorder within the class also benefit from this normalization and evaluate correctly.

Important

The consumer version of Gemini Code Assist on GitHub is being sunset. Starting June 18, 2026, new organization installations will be blocked, and all code review activity will officially cease on July 17, 2026.
For more details on the timeline and next steps, please review the Help Documentation.

[gemini-code-assist]

By normalizing self.actorder in-place rather than using a local variable, we ensure that other references to self.actorder in this class also benefit from the normalization.

Specifically:

1. On line 145, self.is_k_full = (not self.actorder) ... will evaluate correctly. If self.actorder is an Enum member (e.g., ActOrder.NONE), not self.actorder would incorrectly evaluate to False because Python Enum members are always truthy.
2. On line 303, if self.actorder == "group": will correctly match the string value instead of failing when compared against an Enum member or a mixed-case string.

Suggested change

	actorder = getattr(self.actorder, "value", self.actorder)
	if isinstance(actorder, str):
	actorder = actorder.lower()
	load_full_w2 = (
	actorder not in (None, False, "", "static") and self.group_size != -1
	)
	self.actorder = getattr(self.actorder, "value", self.actorder)
	if isinstance(self.actorder, str):
	self.actorder = self.actorder.lower()
	load_full_w2 = (
	self.actorder not in (None, False, "", "static") and self.group_size != -1
	)

[bingxche]

Experiment: disable enable_multithread_load for Kimi-K2.6 (mi325 + mi355x)

Force-pushed this branch to a single recipe-only commit (ac6ae22) that flips enable_multithread_load to false in both K2.6 accuracy recipes — nothing else changed.

Why: the K2.6 _load_w2 narrow IndexError on AMD looks non-deterministic, not a deterministic sharding bug:

• Same load/shard code (_load_w2 narrow unchanged since May 20) and same HF checkpoint (last modified 2026-05-19), yet mi325 failed 6 days straight then passed clean on 6/23 (acc 0.942).
• The failing narrow reports a different TP rank first each day — got 192 / 448 / 192 / 320 / 320 (= 64 × rank for rank 3/7/3/5/5), plus one day it was a segfault in fill_kernel instead. Whoever loses the race raises first → smells like a multithreaded-load race.

This run disables multithreaded loading to see if the IndexError disappears.

Triggered run (mi325 + mi355x kimi-k2.6 only): https://github.com/sgl-project/sglang/actions/runs/28089396987

Note: mi355x has separately been blocked at the hf_xet download stall (7200s timeout) before load even starts, so its result here may still be masked by that.