Add dpa for dsr1 fp4

[faradawn]

Motivation

For better performance, add DP attention for FP4 B200.

Reference SemiAnalysisAI/InferenceX#1792

Modifications

Accuracy Tests

Speed Tests 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 and Merge Process

1. Ping Merge Oncalls to start the process. 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.
   • Common commands include /tag-and-rerun-ci, /tag-run-ci-label, /rerun-failed-ci
4. After green CI and required approvals, ask Merge Oncalls or people with Write permission to merge the PR.

---

CI States

Latest PR Test (Base): ✅ Run #28060588989
Latest PR Test (Extra): ❌ Run #28060588854

[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!

[zijiexia]

This keys on hardware+quantization only, so DP attention is also force-enabled for the low-latency configs (concurrency 4–8), which the prior doc text scoped to high-throughput. Could you confirm against the referenced InferenceX PR (SemiAnalysisAI/InferenceX#1792) whether this recipe was validated for the low-latency scenario, or whether it should be gated on vals.scenario === 'high-throughput'?

[zijiexia]

This special-case diverges from the file's config-driven pattern, where commands render purely from each config's parameters + fieldToFlag. Consider moving these flags into the 4 b200-fp4 config blocks and adding the missing fieldToFlag entries (enable_dp_attention, enable_dp_attention_local_control_broadcast, enable_dp_lm_head) — that also makes per-scenario gating trivial.

[zijiexia]

These emit long-form --tensor-parallel-size / --data-parallel-size, while the rest of the generated command uses short forms (--tp, --ep-size) and fieldToFlag already maps data_parallel_size → 'dp'. Suggest --tp / --dp for copy-paste consistency.

[zijiexia]

3.33 is a calibrated value: when dp-attention is on, server_args.py applies schedule_conservativeness *= 0.3, so 3.33 × 0.3 ≈ 1.0 (the default). Worth a short comment so it isn't later "rounded" to an integer and silently dropped to 0.3 effective.

[zijiexia]

This rewrite drops the previous general --enable-dp-attention --tp 8 --dp 8 hint that also applied to other hardware (e.g. H200 high-throughput). If that's still recommended there, consider keeping a one-line general note alongside the B200-FP4-specific text.