Add DeepseekV3HybridMoeModuleArchitecture

[zhoutong-hai]

DeepSeek V3 uses a hybrid dense + MoE MLP:
- the first 3 layers are dense MLP weights:
    mlp.{gate_proj,up_proj,down_proj}.weight
- the remaining layers are typically MoE (controlled by moe_layer_freq):
    mlp.gate.weight (+ optional e_score_correction_bias),
    mlp.shared_experts.{gate_proj,up_proj,down_proj}.weight,
    mlp.experts.{i}.{gate_proj,up_proj,down_proj}.weight

---

Note

Introduces support for DeepSeek V3’s hybrid dense+MoE layout and wires it into architecture selection.

• Adds DeepseekV3HybridMoeModuleArchitecture with config-driven layer mapping: dense mlp.{gate_proj,up_proj,down_proj} for initial layers, then MoE (mlp.gate[.e_score_correction_bias?], mlp.shared_experts.*, mlp.experts.{i}.*) by moe_layer_freq; includes attention, pre/post weights
• Registers the architecture in architecture/__init__.py (model_type=deepseek_v3)
• Graph runtime: sets arbitrary_types_allowed=True on Task to accept non-pydantic types

^{Written by Cursor Bugbot for commit efce27e. This will update automatically on new commits. Configure here.}

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[zhoutong-hai]

I have read the CLA Document and I hereby sign the CLA

[cursor]

Cursor Bugbot has reviewed your changes and found 1 potential issue.

^{Bugbot Autofix is OFF. To automatically fix reported issues with Cloud Agents, enable Autofix in the Cursor dashboard.}

[cursor]

Missing optional and tied_names for lm_head.weight

High Severity

The lm_head.weight in post_weights is missing optional=True and tied_names=("model.embed_tokens.weight",). Many transformer models (including DeepSeek V3 variants) tie the input embeddings with the output LM head, storing only one copy as model.embed_tokens.weight. Without optional and tied_names, the weight loading will fail when lm_head.weight doesn't exist separately in the checkpoint, as the code has no fallback to look for the weight under its tied name. Other architecture definitions (LLaMA, Mistral) correctly handle this pattern.