[Optimization] default compile rdma, reduce cudagraph buffer size in mm, fix some config bug

[yuanlehome]

Motivation

• PD分离需要依赖rdma_comm包 https://github.com/PaddlePaddle/FastDeploy/tree/develop/fastdeploy/cache_manager/transfer_factory/kvcache_transfer ，但没有默认打开编译且报错不明显。

• 混合并行中cudagraph+tsp会存在hang的情况。

• cudagraph多模组网中申请的hidden_states buffer太大比较占显存。

Modifications

• 默认编译rdma_comm包

  • 仅xpu/gpu

• 修复tsp hang的问题，在开启cudagraph的场景，不捕获capture size < tp_size的shape

• 合理计算buffer shape并修改

Usage or Command

无。

Accuracy Tests

无。

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!

[Copilot]

Pull Request Overview

This PR makes three main optimizations: enables RDMA compilation by default, reduces CUDA graph buffer sizes in vision-language models, and fixes configuration validation logic.

• Enables RDMA extension to compile by default by removing conditional checks on ENABLE_FD_RDMA
• Optimizes memory usage by using max_capture_size instead of max_model_len for CUDA graph buffers
• Fixes RDMA port validation to account for both data parallel and tensor parallel sizes

Reviewed Changes

Copilot reviewed 7 out of 7 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
setup.py	Removes conditional compilation for RDMA extension, making it compile by default
qwen2_5_vl/qwen2_5_vl.py	Reduces buffer size from max_model_len to max_capture_size and optimizes buffer usage with conditional copy
paddleocr_vl/paddleocr_vl.py	Reduces buffer size and simplifies forward logic by removing redundant if-else branches
ernie_vl_rm.py	Reduces buffer size and adds conditional buffer copy for CUDA graph
ernie4_5_vl/ernie4_5_vl_moe.py	Reduces buffer size and adds conditional buffer copy for CUDA graph
engine/args_utils.py	Fixes RDMA port validation to multiply both DP and TP sizes
config.py	Disables sequence parallel MoE when using CUDA graph for mixed/decode roles