[Iluvatar] Support V1_KVCACHE_SCHEDULER and paddleocr-vl rope mode

[wuyujiji]

Motivation

为了适配paddleocr-vl模型，特在天数硬件上支持V1_KVCACHE_SCHEDULER和paddle ocr vl的rope模式。除此之外，还验证了打开V1_KVCACHE_SCHEDULER后，之前适配的ERNIE纯文模型和ERNIE VL模型系列精度均正常

Modifications

Pass

Usage or Command

Pass

Accuracy Tests

Pass

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 10.52632% with 34 lines in your changes missing coverage. Please review.
⚠️ Please upload report for BASE (develop@404cf0e). Learn more about missing BASE report.

Files with missing lines	Patch %	Lines
...executor/layers/attention/iluvatar_attn_backend.py	8.33%	22 Missing ⚠️
...del_executor/models/ernie4_5_vl/ernie4_5_vl_moe.py	20.00%	2 Missing and 2 partials ⚠️
...model_executor/models/paddleocr_vl/paddleocr_vl.py	20.00%	2 Missing and 2 partials ⚠️
fastdeploy/engine/sched/resource_manager_v1.py	0.00%	2 Missing ⚠️
fastdeploy/engine/args_utils.py	0.00%	0 Missing and 1 partial ⚠️
fastdeploy/worker/worker_process.py	0.00%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             develop    #5555   +/-   ##
==========================================
  Coverage           ?   63.80%           
==========================================
  Files              ?      329           
  Lines              ?    41743           
  Branches           ?     6386           
==========================================
  Hits               ?    26636           
  Misses             ?    13081           
  Partials           ?     2026           

Flag	Coverage Δ
GPU	63.80% <10.52%> (?)

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.

[kevincheng2]

天数支持多模请求的多batch嘛？当前v1里边都是放开的，可能要关注一下

[wuyujiji]

天数支持多模请求的多batch嘛？当前v1里边都是放开的，可能要关注一下

@kevincheng2 应该是支持的，有多batch的脚本吗，我可以测一下

[Jiang-Jia-Jun]

LGTM

[Copilot]

Pull request overview

This pull request adds support for V1_KVCACHE_SCHEDULER and PaddleOCR-VL rope mode on Iluvatar hardware. The changes enable the V1 KV cache scheduler on Iluvatar devices and implement a new rope mode specifically for PaddleOCR-VL models while maintaining backward compatibility with ERNIE text and VL model series.

Key Changes:

• Refactored timeout mechanism in tests using signal-based approach
• Updated dependency versions (paddleformers 0.4.0, paddle packages to dev20251103/20251107)
• Extended V1_KVCACHE_SCHEDULER support to Iluvatar platform
• Modified rope embedding handling to support both interleaved and non-interleaved modes
• Added new custom operators for V1 scheduler support (update_inputs_v1, recover_decode_task, get_img_boundaries)

Reviewed changes

Copilot reviewed 24 out of 24 changed files in this pull request and generated 12 comments.

Show a summary per file

File	Description
tests/ci_use/iluvatar_UT/utils.py	New utility module with signal-based timeout decorator
tests/ci_use/iluvatar_UT/*.py	Refactored tests to use centralized timeout utility, updated expected outputs
tests/ci_use/iluvatar_UT/bench_gsm8k.py	New benchmark script for GSM8K dataset evaluation
scripts/run_ci_iluvatar.sh	Improved CI script with better error logging
requirements_iluvatar.txt	Updated paddleformers version to 0.4.0
fastdeploy/worker/*.py	Enabled V1 scheduler for Iluvatar, adjusted rope embedding logic
fastdeploy/model_executor/layers/attention/iluvatar_attn_backend.py	Major refactoring of rope embedding handling for batch processing
fastdeploy/model_executor/ops/iluvatar/paged_attention.py	Added rope_batch_stride and is_interleaved_rope_mode parameters
fastdeploy/model_executor/models//.py	Added transpose operations for mixed attention mode
custom_ops/setup_ops.py	Added new operator source files to build
custom_ops/iluvatar_ops/*.cu	Updated attention kernels with rope mode support and batch stride
custom_ops/gpu_ops/get_padding_offset.cu	Fixed warp size for Iluvatar (64 vs 32)
docs/**/*.md	Extensive documentation updates for Iluvatar setup and model deployment
.github/workflows/ci_iluvatar.yml	Updated Docker image and runner configuration