[PD Disaggregation][XPU] Add XPU support for PD disaggregation

[ddchenhao66]

Motivation

XPU支持PD分离

Modifications

• 增加init_signal_layerwise算子、修改open_shm_and_get_meta_signal算子接口
• 增加get_peer_mem_addr算子，适配xpu RDMA地址空间注册
• xpu_model_runner修改v0调度的插入逻辑，适配PD分离的情形；增加xpu create_kv_signal_sender/destroy_kv_signal_sender算子，实现异步写共享变量
• prefix_cache_manager起cache进程时强制重置XPU_VISIBLE_DEVICES，解决非0卡起服务报错问题

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 sign our Contributor License Agreement before we can accept your contribution.

---

ddchenhao66 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.}

[Copilot]

Pull Request Overview

This PR adds XPU hardware support for PD (Prefill-Decode) disaggregation feature, which allows splitting prefill and decode phases across different compute nodes. The implementation mirrors existing CUDA/GPU support by adding XPU-specific code paths throughout the attention, cache management, and worker execution layers.

• Extends attention backends to support XPU with PD disaggregation modes ("per_chunk" and "per_query")
• Adds XPU signal handling and inter-process communication for KV cache coordination
• Refactors platform-specific operation imports to support both CUDA and XPU through a unified interface
• Updates cache manager to handle XPU-specific memory addressing and device visibility

Reviewed Changes

Copilot reviewed 16 out of 16 changed files in this pull request and generated 12 comments.

Show a summary per file

File	Description
fastdeploy/worker/xpu_model_runner.py	Adds PD disaggregation mode support in XPU worker, including decode node handling and KV signal sender lifecycle management
fastdeploy/model_executor/layers/attention/xpu_attn_backend.py	Implements XPU attention backend with PD disaggregation initialization and signal handling
fastdeploy/model_executor/layers/attention/utils.py	Adds XPU device visibility environment variable support (XPU_VISIBLE_DEVICES)
fastdeploy/model_executor/forward_meta.py	Adds kv_signal_sender field to XPUForwardMeta for PD disaggregation
fastdeploy/model_executor/layers/attention/ops/*.py	Adds XPU platform branches to signal operation wrappers
fastdeploy/cache_manager/*.py	Refactors ops imports to platform-agnostic interface and adds XPU memory address handling
custom_ops/xpu_ops/src/ops/*.cc	Implements XPU-specific C++ operations for signal handling and shared memory

[hong19860320]

LGTM