Expose Model Parallelism Information

[JD-ETH]

Motivation

We want to enable model shards to be initialized outside of sglang in a way that's identical to sglang's ModelLoader.
The external initialization workflow will look something like this:

sglang.srt.server_args._global_server_args = server_args
model_parallelism_info = engine.get_parallelism_config(rank) 
with ParallelismContext(RankParallelismConfig.from_dict(model_parallelism_info)):
   model = get_model(
                model_config=model_config,
                load_config=load_config,
                device_config=device_config,
            )

The primary use case of this is to enable train -> inference weight transfer. with 14997, we already have transfer engine and weight registration, but the transfer can only happen between weights of identical shape. One option is to instantiate the model shards outside, register the weights, and send them to sglang via the existing registrations.

I will however leave draft_tp for future work. The test currently verifies: TP, EP and Moe with DP Attention, with the CI test only has the most simple tp case.

Modifications

Whenever transfer engine is enabled, we additionally expose the full parallelism configurations of tp/pp/ep/attn_tp/attn_dp information through http API. The info is propagated to the _global_states just like the rdma weight register does at initialization time, so no performance impact.

A small fix was made to pass dp_parallel_controller's scheduler_infos back to the global engine instance. This allows remote_weight_info APIs to work with dp>1.

Accuracy Tests

python -m sglang.launch_server --model-path qwen/qwen2.5-0.5b-instruct --remote-instance-weight-loader-start-seed-via-transfer-engine
curl http://localhost:30000/get_parallelism_config
returns:
{"rank":0,"tp_size":1,"tp_rank":0,"pp_size":1,"pp_rank":0,"ep_size":1,"ep_rank":0,"attn_tp_size":1,"attn_tp_rank":0,"attn_dp_size":1,"attn_dp_rank":0,"world_size":1,"global_rank":0} etc

the test_parallelism_context_integration.py shows the intended use case and that the model parameters match.

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.

Review Process

1. Ping Merge Oncalls to start the PR flow. See the PR Merge Process.
2. Get approvals from CODEOWNERS and other reviewers.
3. Trigger CI tests with comments or contact authorized users to do so.
   • /tag-run-ci-label, /rerun-failed-ci, /tag-and-rerun-ci
4. After green CI and required approvals, ask Merge Oncalls to merge.

[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!

[stmatengss]

/tag-and-rerun-ci

[JD-ETH]

/rerun-failed-ci

[zhaochenyang20]

1. add the commands of HTTPS;
2. modify sglang document;

[zhaochenyang20]

/rerun-failed-ci

[JD-ETH]

add to run suite

[JD-ETH]

refactored to new PR without all the piping: #20907