[Fridge003]

Recreates #39275 after its branch was accidentally deleted. Restores the same head commit (f33320f18a) and code diff.

Motivation

Test upgrading sgl-deep-gemm from 0.1.7 to 0.2.0rc0 while preserving SGLang batch-invariant outputs. DeepGEMM 0.2 defaults BF16 GEMM to cuBLASLt, whose reduction can change with batch size; the batch-invariant override must explicitly select the deterministic kernel.

Modifications

• Bump sgl-deep-gemm to 0.2.0rc0.
• Select DeepGEMM's deterministic BF16 kernel within the batch-invariant call and restore the previous setting in finally; retain compatibility with older wheels lacking the mode API.
• Add a random BF16 regression test across batch sizes with zero tolerance and verify mode preservation.

Accuracy Tests

Controlled GPU validation used the original PR #39275 head (c6c4e226f2) with the same fix, holding model snapshots, input caches, and other packages fixed.

Test	0.1.7	0.2.0rc0	0.2.0rc0 + fix
H200 FA3 prefix tokens/logprobs	Pass	Fail	Pass (145 responses identical)
2xH200 DP breakable prefill/decode KL	0 / 0	5.23e-4 / 1.089e-3	0 / 0
4xB200 Inkling NVFP4 + DSPARK input/output KL	0	9.94e-4	0
Random BF16 batch-invariance regression, H200/B200	Pass in legacy probe	Fail	Pass

All exact-equality/KL thresholds are unchanged. The full B200 batch-invariant unit file passes all 7 tests. The migrated patch and both resulting files are identical to the validated fix; Python 3.12 pre-commit checks pass after migration.

Speed Tests and Profiling

A focused same-B200 check (BF16 N=K=4096, M=1/16/64/256) measured eager latency at 14.39–14.77 µs on 0.1.7 versus 14.51–14.89 µs with the fix; CUDA-graph replay was 14.33–14.35 µs versus 10.25–12.29 µs. These four shapes are not a full serving benchmark.

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 #34896686768
Latest PR Test (Extra): ❌ Run #34896686877
Latest PR Test (AMD ROCm 10): ❌ Run #34896686574

[Fridge003]

Test runs:
https://github.com/sgl-project/sglang/actions/runs/34788188380
https://github.com/sgl-project/sglang/actions/runs/34788188290