[Feature] Support FP4 communication quantization

[lizexu123]

Motivation

支持fp4 通信量化,以hidden_size = 7168为例子

路径	数据	Scale	总计 (per token)
BF16	7168 × 2B = 14336B	无	14336B
FP4	3584 × 1B = 3584B	112 × 4B = 448B	4032B

FP4 通信量 = BF16 的 28%，减少约 3.5 倍。在 EP 场景下（8卡 all-to-all），通信往往是瓶颈

💡 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!

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you all sign our Contributor License Agreement before we can accept your contribution.
1 out of 2 committers have signed the CLA.

✅ lizexu123
❌ root

---

root seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}

[PaddlePaddle-bot]

🤖 AI Code Review | 2026-04-19 23:37 CST\n\n## 📋 Review 摘要\n\nPR 概述：支持 FP4 通信量化，在 EP prefill dispatch 场景下将通信量降低至 BF16 的约 28%，显著减少 all-to-all 通信瓶颈。\n变更范围：custom_ops/gpu_ops/moe/（C++ 算子）、model_executor/layers/quantization/nvfp4.py（Python 量化逻辑）、envs.py（环境变量）、forward_meta.py、utils.py\n影响面 Tag：OP Quantization\n\n### 📝 PR 规范检查\n\nPR 描述中 Modifications、Usage or Command、Accuracy Tests 三个章节均为空，建议补充。该 PR 影响模型前向计算（MoE FFN 路径），应提供精度测试结果。\n\n描述模板（可直接复制）：\n\n## Modifications\n1. 新增环境变量 FD_USE_NVFP4_COMM_QUANT 控制 FP4 通信量化开关\n2. 在 nvfp4.py apply_ep_prefill 中增加 FP4 预量化 dispatch 路径\n3. C++ 算子 PrefillPermuteToMaskedGemm 新增 UINT8 数据类型支持\n4. utils.py 增加未初始化权重的防御性检查\n\n## Usage or Command\nFD_USE_NVFP4_COMM_QUANT=1 python -m fastdeploy.entrypoints.openai.api_server ...\n\n## Accuracy Tests\n（请补充 EP 场景下开启/关闭 FP4 通信量化的精度对比结果）\n\n\n### 问题\n\n| 级别 | 文件 | 概述 |\n|------|------|------|\n| 🟡 建议 | utils.py:136 | 调试日志使用非正式中文，不适合生产代码 |\n| ❓ 疑问 | forward_meta.py:156 | audio_token_num 字段与本 PR 功能无关且未被使用 |\n\n### 总体评价\n\nFP4 通信量化的核心逻辑（预量化 → dispatch → scale swizzle → masked GEMM）实现合理，注释清晰，FP4/BF16 两条路径分支清楚。建议补充 PR 描述中的精度测试结果，并清理两处小问题。"

[PaddlePaddle-bot]

🟡 建议 调试日志使用了非正式中文 "权重没初始化啊！"，不适合生产代码。

1. 日志语言应与项目其他日志保持一致（英文），且语气应正式；
2. 日志级别建议使用 logger.warning 而非 logger.info——权重未初始化导致跳过 transpose 可能掩盖上游问题，warning 级别更有助于排查。

建议修改为：

if not weight._is_initialized():
    logger.warning("Weight '%s' is not initialized, skipping transpose.", weight_name)
    return

[PaddlePaddle-bot]

❓ 疑问 audio_token_num 字段与本 PR 的 FP4 通信量化功能似乎无关，且搜索整个代码库后未发现任何读取或写入此字段的代码。

请确认：这个字段是否应该在单独的 PR 中提交？如果是后续功能的前置准备，建议在注释中说明用途。