[Graph Optimization][CINN] Use CINN in PaddleOCR-VL ViT part

[SigureMo]

Motivation

使用 CINN 加速 PaddleOCR ViT 部分，对其中 _run_encoder_layer 装饰 to_static

Modifications

• ViT encoder 拆分两部分，循环部分用 to_static 装饰
• 添加为 ViT encoder 的 warmup 逻辑，目前仅针对 PaddleOCR-VL 模型，后续再考虑其他模型，会考虑将逻辑抽象一下
• 原 tests/e2e/test_paddleocr_vl_serving.py 同时测动态图+CINN

Usage or Command

目前开启 compile 与 LLM 部分方式相同：

• graph_opt_level=0 纯动态图，走 fused ops
• graph_opt_level=1 SOT 动转静，走 native ops，性能有所下降
• graph_opt_level=2 CINN，走 native ops + CINN fusion，性能与 fused ops 持平且略有超越

python -m fastdeploy.entrypoints.openai.api_server \
        --model /root/paddlejob/tmpspace/MODELS/PaddlePaddle/PaddleOCR-VL \
        --port 8295 \
        --metrics-port 8296 \
        --engine-worker-queue-port 8297 \
        --cache-queue-port 55660 \
        --max-model-len 16384 \
        --max-num-batched-tokens 16384 \
        --gpu-memory-utilization 0.7 \
        --max-num-seqs 256 \
        --workers 2 \
        --graph-optimization-config '{"graph_opt_level":2, "use_cudagraph":true}'

	动态图无 fused（tokens/s）	动态图手工 fused（tokens/s）	CINN（tokens/s）
HZZ1	15394.50	16944.64 (+10%)	17211.10 (+12%)
A10	4743.69	6359.45 (+34%)	6509.28（+37%)

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!

[codecov-commenter]

Codecov Report

❌ Patch coverage is 93.47826% with 3 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@5d9b5e4). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
fastdeploy/worker/gpu_model_runner.py	93.10%	1 Missing and 1 partial ⚠️
fastdeploy/worker/model_runner_base.py	50.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #5223   +/-   ##
==========================================
  Coverage           ?   60.04%           
==========================================
  Files              ?      329           
  Lines              ?    40980           
  Branches           ?     6210           
==========================================
  Hits               ?    24606           
  Misses             ?    14517           
  Partials           ?     1857           

Flag	Coverage Δ
GPU	60.04% <93.47%> (?)

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull request overview

本 PR 为 PaddleOCR-VL 模型的 ViT 编码器部分引入了 CINN 编译支持，通过图优化提升推理性能。主要实现了三级图优化策略：level 0（纯动态图+fused ops）、level 1（SOT 动转静+native ops）、level 2（CINN 编译+native ops），其中 CINN 模式在性能测试中相比动态图有 12-37% 的提升。

• 将 SiglipEncoder 的循环部分提取为 _run_encoder_layer 方法，支持 to_static 装饰和 CINN 编译
• 为 vision encoder 添加 warmup 逻辑和编译流程，目前仅支持 PaddleOCR-VL 模型
• 扩展端到端测试以覆盖动态图和 CINN 两种模式

Reviewed changes

Copilot reviewed 8 out of 8 changed files in this pull request and generated 12 comments.

Show a summary per file

File	Description
tests/e2e/test_paddleocr_vl_serving.py	参数化测试以覆盖 graph_opt_level 0 和 2 两种配置
fastdeploy/worker/model_runner_base.py	添加 vision_encoder_compile 基类方法
fastdeploy/worker/gpu_worker.py	在 warm-up 流程中集成 vision encoder 编译
fastdeploy/worker/gpu_model_runner.py	实现 vision encoder 的 CINN 编译和 warmup 逻辑
fastdeploy/model_executor/models/paddleocr_vl/siglip_ops.py	添加动态形状处理和统一标记以支持静态图编译
fastdeploy/model_executor/models/paddleocr_vl/siglip.py	重构编码器，提取 _run_encoder_layer 用于 CINN 编译
fastdeploy/engine/engine.py	添加 SOT_ENABLE_COMPILE_TIME_LIMIT 环境变量
fastdeploy/engine/async_llm.py	添加 SOT_ENABLE_COMPILE_TIME_LIMIT 环境变量

[gongshaotian]

LGTM

[ming1753]

LGTM