refactor context parallel state

[dongjiyingdjy]

Motivation

Context parallelism is essential in long context LLM inference. It splits a long input sequence across multiple GPUs so attention can be computed in parallel, drastically reducing latency, which enables practical million-token context windows.
Previously, DeepSeek-V3.2 already supported CP and could use it together with DP. We aim to support the combination of CP, DP, and TP, and make it easier to apply to other models. To achieve this, we first refactored the original implementation.

Modifications

• Add _ATTN_CP, _ATTN_TP and _MOE_CP as new GroupCoordinator;
• Add attn_cp_size and moe_cp_size as new server_agrs;
• Updated the relevant interfaces and processing logic in Scheduler.

Accuracy Tests

============================================================
DeepSeek-V3.2-Exp CP Single Node Results Summary
Dataset: gsm8k
Baseline: 0.935
============================================================

Model 1: deepseek-ai/DeepSeek-V3.2-Exp
  Accuracy: PASS
  Score: 0.980

Model 2: deepseek-ai/DeepSeek-V3.2-Exp
  Accuracy: PASS
  Score: 0.980

============================================================
OVERALL: ALL TESTS PASSED
============================================================

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.

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]

Summary of Changes

Hello @dongjiyingdjy, I'm Gemini Code Assist¹! I'm currently reviewing this pull request and will post my feedback shortly. In the meantime, here's a summary to help you and other reviewers quickly get up to speed!

This pull request refactors the context parallel state to support attention and MoE context parallelism. It introduces new group coordinators and updates initialization functions to manage these parallel processing capabilities. The changes also include modifications to scheduler processes and the addition of new server arguments for configuration.

Highlights

• Context Parallelism Refactor: This PR introduces context parallelism for attention and MoE layers, enhancing parallel processing capabilities.
• Group Coordinator Updates: New group coordinators (_ATTN_TP, _ATTN_CP, _MOE_CP) are added to manage attention and MoE context parallelism.
• Initialization Function Modification: The initialize_model_parallel function is updated to include parameters for attention and MoE context parallel sizes.
• Scheduler Processes: Scheduler processes are modified to pass attention and MoE context parallel ranks.
• Server Arguments: New server arguments attn_cp_size and moe_cp_size are added to configure context parallelism.

🧠 New Feature in Public Preview: You can now enable Memory to help Gemini Code Assist learn from your team's feedback. This makes future code reviews more consistent and personalized to your project's style. Click here to enable Memory in your admin console.

Using Gemini Code Assist

The full guide for Gemini Code Assist can be found on our documentation page, here are some quick tips.

Invoking Gemini

You can request assistance from Gemini at any point by creating a comment using either /gemini <command> or @gemini-code-assist <command>. Below is a summary of the supported commands on the current page.

Feature	Command	Description
Code Review	/gemini review	Performs a code review for the current pull request in its current state.
Pull Request Summary	/gemini summary	Provides a summary of the current pull request in its current state.
Comment	@gemini-code-assist	Responds in comments when explicitly tagged, both in pull request comments and review comments.
Help	/gemini help	Displays a list of available commands.

Customization

To customize Gemini Code Assist for GitHub experience, repository maintainers can create a configuration file and/or provide a custom code review style guide (such as PEP-8 for Python) by creating and adding files to a .gemini/ folder in the base of the repository. Detailed instructions can be found here.

Limitations & Feedback

Gemini Code Assist may make mistakes. Please leave feedback on any instances where its feedback is incorrect or counter productive. You can react with 👍 and 👎 on @gemini-code-assist comments. If you're interested in giving your feedback about your experience with Gemini Code Assist for Github and other Google products, sign up here.

You can also get AI-powered code generation, chat, as well as code reviews directly in the IDE at no cost with the Gemini Code Assist IDE Extension.

Footnotes

1. Review the Privacy Notices, Generative AI Prohibited Use Policy, Terms of Service, and learn how to configure Gemini Code Assist in GitHub here. Gemini can make mistakes, so double check it and use code with caution. ↩

[gemini-code-assist]

Code Review

This pull request introduces context parallelism for attention and MoE layers, which is a significant refactoring. The changes are extensive, touching many files to plumb through the new configuration and rank information. The core logic for creating the new parallel groups has been added.

My review focuses on the correctness of the new parallelism group initialization. While the logic for attention context parallelism (_ATTN_CP) and other groups seems correct, I've found a critical issue in the initialization of the MoE context parallel group (_MOE_CP). The current implementation appears to create groups across pipeline stages instead of within a single stage, which is incorrect for context parallelism.

I've provided a detailed comment with a suggested fix for this issue. Please address this to ensure the correctness of MoE context parallelism.

Also, there is a small typo in the pull request title: "refatcor" should be "refactor".

[gemini-code-assist]

The logic for creating the _MOE_CP (MoE Context Parallel) group appears to be incorrect. It seems to be creating groups across pipeline parallel stages, similar to how pipeline parallel groups are formed, rather than creating context parallel groups within a single pipeline stage.

A context parallel group for MoE should group ranks that handle different parts of the context but the same expert and tensor slice. The current implementation:

for i in range(num_tensor_model_parallel_groups):
    for j in range(moe_tp_size * moe_ep_size):
        st = i * tensor_model_parallel_size + j
        en = (i + 1) * tensor_model_parallel_size + j
        ranks = list(range(st, en, moe_tp_size * moe_ep_size))
        group_ranks.append(ranks)

Here, i iterates through pipeline stages, and en points to a rank in the next pipeline stage, which is incorrect for a context parallel group.

A correct implementation should iterate within a single pipeline stage. Assuming a rank layout of (cp, ep, tp) within a tensor parallel group, the logic should be something like this:

        for i in range(num_tensor_model_parallel_groups):
            for j in range(moe_ep_size):
                for k in range(moe_tp_size):
                    # Assuming a rank layout of (cp, ep, tp)
                    base = i * tensor_model_parallel_size + j * moe_tp_size + k
                    stride = moe_ep_size * moe_tp_size
                    ranks = [base + c * stride for c in range(moe_cp_size)]
                    group_ranks.append(ranks)

[Fridge003]

Please make sure this PR can pass this unit test
https://github.com/sgl-project/sglang/blob/main/test/registered/8-gpu-models/test_deepseek_v32_cp_single_node.py

[dongjiyingdjy]

Please make sure this PR can pass this unit test https://github.com/sgl-project/sglang/blob/main/test/registered/8-gpu-models/test_deepseek_v32_cp_single_node.py

Both tests already passed. Thanks!

[Fridge003]

/tag-and-rerun-ci

[Fridge003]

Wait for CIs

[llc-kc]

This PR disable PP+CP, will this be supported in the future?