Support PP for zmq_to_scheduler

[gty111]

Motivation

Follow up PRs for #12263 to support PP for zmq_to_scheduler

Modifications

Accuracy Tests

Benchmarking and Profiling

Checklist

• Format your code according to the Format code with pre-commit.
• Add unit tests according to the Run and add unit tests.
• Update documentation according to Write documentations.
• Provide accuracy and speed benchmark results according to Test the accuracy and Benchmark the speed.
• Follow the SGLang code style guidance.
• Work with maintainers to merge your PR. See the PR Merge Process

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[Copilot]

Pull request overview

This PR enables Pipeline Parallelism (PP) support for the zmq_to_scheduler encoder transfer backend in encoder-decoder disaggregated inference scenarios. The key insight is that MM (multimodal) processing only needs to occur at the first PP stage (pp_rank 0), with subsequent stages receiving pre-processed requests from upstream stages.

Key changes:

• Restricted MM receiver initialization and processing to pp_rank 0 only
• Changed synchronization scope from world_size (all ranks) to tp_size (TP ranks within a PP stage)
• Removed the embedding_ports mechanism in favor of direct encoder-to-scheduler communication per TP group

Reviewed changes

Copilot reviewed 6 out of 6 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
python/sglang/srt/server_args.py	Removed validation that prevented PP when using zmq_to_scheduler
python/sglang/srt/models/qwen2_5_vl.py	Simplified weight loading to skip missing weights consistently across modes
python/sglang/srt/managers/io_struct.py	Removed embedding_ports field from request input structures
python/sglang/srt/managers/tokenizer_manager.py	Removed embedding_ports parameter from tokenized object creation
python/sglang/srt/managers/scheduler.py	Added pp_rank checks for MM receiver initialization and processing; added tp_group parameter
python/sglang/srt/disaggregation/encode_receiver.py	Changed synchronization to TP-only scope, improved device placement, removed embedding_port logic

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[ShangmingCai]

/tag-and-rerun-ci

[ShangmingCai]

LGTM, let's wait for the CI.

[ShangmingCai]

/rerun-failed-ci

[ShangmingCai]

LGTM. The modification of the scheduler part is safe since it is all EPD-related code.

Full green CI: https://github.com/sgl-project/sglang/actions/runs/20449993152/job/58772672138?pr=15312