[FDConfig] Unify num_experts_per_tok to moe_k in ModelConfig for MoE model compatibility

[xyxinyang]

Motivation

💡 If this PR is a Cherry Pick, the PR title needs to follow the format by adding the [Cherry-Pick] label at the very beginning and appending the original PR ID at the end. For example, [Cherry-Pick][CI] Add check trigger and logic(#5191)

💡 如若此PR是Cherry Pick，PR标题需遵循格式，在最开始加上[Cherry-Pick]标签，以及最后面加上原PR ID，例如[Cherry-Pick][CI] Add check trigger and logic(#5191)

Unify num_experts_per_tok to moe_k in ModelConfig for MoE model compatibility. This enables R3 support for models like Qwen3VLMOE, DeepSeek V3, and other MoE models that use num_experts_per_tok instead of moe_k in their config.

Modifications

• Added unified field mapping in ModelConfig (fastdeploy/config.py)

Usage or Command

N/A

Accuracy Tests

N/A

Checklist

• Add at least a tag in the PR title.
  • Tag list: [[FDConfig],[APIServer],[Engine], [Scheduler], [PD Disaggregation], [Executor], [Graph Optimization], [Speculative Decoding], [RL], [Models], [Quantization], [Loader], [OP], [KVCache], [DataProcessor], [BugFix], [Docs], [CI], [Optimization], [Feature], [Benchmark], [Others], [XPU], [HPU], [GCU], [DCU], [Iluvatar], [Metax]]
  • You can add new tags based on the PR content, but the semantics must be clear.
• Format your code, run pre-commit before commit.
• Add unit tests. Please write the reason in this PR if no unit tests.
• Provide accuracy results.
• If the current PR is submitting to the release branch, make sure the PR has been submitted to the develop branch, then cherry-pick it to the release branch with the [Cherry-Pick] PR tag.

[paddle-bot]

Thanks for your contribution!

[PaddlePaddle-bot]

🤖 AI Code Review | 2026-04-20 14:57 CST

📋 Review 摘要

PR 概述：在 ModelConfig.override_name_from_config() 中新增 num_experts_per_tok → moe_k 的字段映射，以支持 Qwen3VLMOE、DeepSeek V3 等使用 num_experts_per_tok 的 MoE 模型。
变更范围：fastdeploy/config.py — Config 字段映射
影响面 Tag：FDConfig

问题

级别	文件	概述
🟡 建议	多个模型文件	统一映射不完整，下游消费方仍直接使用 num_experts_per_tok

---

🟡 建议：统一映射后，下游消费方未同步切换到 moe_k

本次 PR 在 override_name_from_config() 中将 num_experts_per_tok 映射为 moe_k，映射逻辑本身是正确的。但代码库中仍有多处直接使用 fd_config.model_config.num_experts_per_tok 而非统一后的 moe_k：

文件	行号	当前用法
model_executor/models/qwen3moe.py	69	top_k=fd_config.model_config.num_experts_per_tok
model_executor/models/deepseek_v3.py	167	top_k=fd_config.model_config.num_experts_per_tok
model_executor/models/gpt_oss.py	120	top_k=fd_config.model_config.num_experts_per_tok
model_executor/models/glm4_moe.py	176	top_k=fd_config.model_config.num_experts_per_tok
model_executor/layers/moe/routing_indices_cache.py	170	对 Glm4MoeForCausalLM 特殊分支使用 num_experts_per_tok

建议：

1. 将上述文件中的 fd_config.model_config.num_experts_per_tok 统一替换为 fd_config.model_config.moe_k，使 "Unify" 目标真正达成。
2. routing_indices_cache.py:169-172 中对 Glm4MoeForCausalLM 的硬编码特殊分支在统一映射后可以移除，直接使用 moe_k，简化逻辑：

# 当前（可简化）：
if fd_config.model_config.architectures[0] == "Glm4MoeForCausalLM":
    self.moe_top_k = fd_config.model_config.num_experts_per_tok
else:
    self.moe_top_k = fd_config.model_config.moe_k

# 建议改为：
self.moe_top_k = fd_config.model_config.moe_k

总体评价

映射逻辑正确，不会引入 bug。但作为一个标题为 "Unify" 的 PR，建议一并更新下游消费方以完成完整的统一，否则后续维护者可能不清楚应该使用 moe_k 还是 num_experts_per_tok。

[gongshaotian]

LGTM