Fix incorrect attribute mapping relationships in GLM MoE DSA Config

[Dovis01]

Fix GlmMoeDsaConfig legacy head_dim overwriting qk_rope_head_dim

Summary

Loading the same GLM-5 config.json yields different attention dimensions between transformers v5.3.0 and v5.4.0. v5.4.0 silently corrupts qk_rope_head_dim due to a new attribute_map entry.

v5.3.0 vs v5.4.0 and after

For example: Checkpoint config.json:

"head_dim": 192,
"qk_nope_head_dim": 192,
"qk_rope_head_dim": 64

	v5.3.0	v5.4.0
Config class	Custom __init__	@strict dataclass, inherits Glm4MoeLiteConfig
attribute_map	No head_dim mapping	"head_dim": "qk_rope_head_dim" (inherited)
qk_rope_head_dim after load	64 ✓	192 ✗ (overwritten by legacy head_dim)
head_dim in config output	192 (kept as separate legacy field)	absent (aliased into qk_rope_head_dim)
qk_head_dim	256 (192+64) ✓	384 (192+192) ✗

Why v5.3.0 works

Custom __init__ sets qk_rope_head_dim from the explicit JSON field first; legacy head_dim=192 is stored separately and does not overwrite qk_rope_head_dim.

Why v5.4.0 breaks

1. qk_rope_head_dim=64 is set as a dataclass field
2. Legacy head_dim=192 arrives later in __post_init__
3. attribute_map redirects head_dim → qk_rope_head_dim, overwriting 64 with 192

Downstream impact (v5.4.0 only)

• MLA attention projection shapes use wrong rope dim (192 vs 64)
• Indexer Q/K split along rope dim is wrong
• RoPE inv_freq length is wrong (getattr(config, "head_dim") also resolves to 192)
• Inference engines reading config.qk_rope_head_dim (e.g. SGLang NSA/MLA backends) get incorrect KV cache / RoPE layout

Fix

1. Override attribute_map in modular_glm_moe_dsa.py — drop "head_dim": "qk_rope_head_dim" for GlmMoeDsaConfig

[ArthurZucker]

39f751a is what changed it, we might have to fix a test maybe? otherwise makes sense

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[github-actions]

[For maintainers] Suggested jobs to run (before merge)

run-slow: glm_moe_dsa

[ArthurZucker]

As its urgent fine by me!