[yueming-yuan]

Loading EAGLE draft weights for a DeepSeek-V3 NextN model through the unified weight loading API crashes on missing attributes.

DeepseekV3ForCausalLMNextN.__init__ calls nn.Module.__init__ and rebuilds the parts of the parent init it needs, so it never reaches where DeepseekV3ForCausalLM.__init__ sets fuse_qkv_a_proj, stacked_params_mapping and expert_params_mapping — the three attributes the unified loader reads off the model.

Rather than copy that block into the subclass, this extracts it into DeepseekV3ForCausalLM.init_unified_loader_mappings() and calls it from both __init__s, right after determine_num_fused_shared_experts() (which expert_params_mapping depends on). A subclass that builds its own __init__ now has one call to make instead of a block to keep in sync.

fuse_qkv_a_proj is still assigned early in the parent, before packed_modules_mapping needs it; the helper recomputes the same value from config, so the parent behaviour is unchanged.

Repro: GLM-5.2 744B with --speculative-algorithm EAGLE on the glm_moe_dsa (DeepSeek-V3 lineage) path.

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ❌ Run #30877585251
Latest PR Test (Extra): ❌ Run #30877585103

[gemini-code-assist]

Caution

The consumer version of Gemini Code Assist on GitHub has been sunset. All code review activity has officially ceased.