Fix score calculation for MOE Learable

[chang-wenbin]

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)

Modifications

Usage or Command

Accuracy Tests

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]

🤖 Paddle-CI-Agent | pr_review | 2026-05-24 17:20:14

📋 Review 摘要

PR 概述：将 MOE Learnable 路由中 routed_scaling_factor_learnable 的 per-expert 缩放操作从 moe_expert_dispatch 调用前移至调用后，修复缩放结果被覆盖导致无效的 Bug。

变更范围：fastdeploy/model_executor/layers/moe/fused_moe_cutlass_backend.py（cutlass MoE backend，apply_tp 函数）

影响面 Tag：[OP]

问题

级别	文件	概述
🟡 建议	fused_moe_cutlass_backend.py:483	其他 MoE backend 可能存在相同 routed_scaling_factor_learnable 缺陷，建议同步排查
📝 PR 规范	—	标题缺少官方 Tag；Motivation / Modifications / Accuracy Tests 段落均为空

📝 PR 规范检查

标题缺少官方 Tag（当前 "Fix score calculation for MOE Learable" 无 [Tag] 前缀，且 "Learable" 存在拼写错误）；PR 描述中 Motivation、Modifications、Usage or Command、Accuracy Tests 等 section 均为空或仅含占位注释，不符合模板要求。

标题建议（可直接复制）：

• [BugFix] Fix score calculation for MOE Learnable routing

PR 描述建议（可直接复制，必须复刻 checklist §D2 模板的完整结构）：

## Motivation
`routed_scaling_factor_learnable` 为 True 时，原代码在 `moe_expert_dispatch` 调用**之前**对 `topk_weights` 乘以 per-expert scale；而 `moe_expert_dispatch(topk_only_mode=True)` 会在内部重新计算并返回新的 `topk_weights`，导致之前的缩放结果被覆盖、缩放完全无效，最终造成 MOE Learnable 场景下 score 计算错误。

## Modifications
- `fastdeploy/model_executor/layers/moe/fused_moe_cutlass_backend.py`：将 `routed_scaling_factor_learnable` 的 per-expert 缩放逻辑从 `moe_expert_dispatch` 调用前移至调用后，确保缩放作用于最终使用的 `topk_weights`；同时将 `get_moe_scores` 返回的 `topk_weights`/`topk_idx` 改为丢弃（`_`, `__`），因为这两个值会被 `moe_expert_dispatch` 的返回值覆盖。

## Usage or Command
N/A

## Accuracy Tests
N/A

## Checklist

- [x] 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.
- [x] 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.

总体评价

修复逻辑正确：原代码在 moe_expert_dispatch 会覆盖 topk_weights 的情况下提前对其应用 learnable 缩放，导致缩放无效；本次将缩放移到 dispatch 之后，确保最终实际使用的权重得到正确缩放。建议同步排查其他 MoE backend（deepgemm / triton / marlin / wint2 / blackwell）中是否存在相同缺陷。

[PaddlePaddle-bot]

🟡 建议 请同步排查其他 MoE backend（fused_moe_deepgemm_backend.py、fused_moe_triton_backend.py、fused_moe_marlin_backend.py、fused_moe_wint2_backend.py、fused_moe_blackwell_backend.py）中是否存在相同的 routed_scaling_factor_learnable 缺陷（即在 moe_expert_dispatch 调用前对 topk_weights 应用缩放，而该调用会覆盖 topk_weights）。

修复策略：在各 backend 中同样将缩放逻辑移到 moe_expert_dispatch 返回之后。