[AMD] Fix AttributeError in GeneratedSharedPrefixDataset.from_args for in-process callers

[michaelzhang-ai]

Restore-only fix — no new logic and no behavior change. It reverts two fields back to the pre-existing getattr(args, ..., default) convention used everywhere else in this file; see Modifications.

Motivation

The 1-GPU nightly tests fail in test/registered/bench_fn/test_bench_serving_functionality.py::test_gsp_multi_turn with:

File ".../sglang/benchmark/datasets/generated_shared_prefix.py", line 58, in from_args
    group_distribution = args.gsp_group_distribution
AttributeError: 'types.SimpleNamespace' object has no attribute 'gsp_group_distribution'

GeneratedSharedPrefixDataset.from_args reads args.gsp_group_distribution and args.gsp_zipf_alpha via direct attribute access. These options have argparse defaults ("uniform" / None) on the bench_serving.py CLI, but in-process callers that build a SimpleNamespace through get_benchmark_args (e.g. test_gsp_multi_turn) never set them, so the access raises before the method's own defensive validation can run. Every other optional gsp_* field in from_args already uses getattr(args, ..., default), and the method's comment states it is meant to tolerate hand-built Namespaces.

This is shared benchmark-client code and the test is registered for both suites (register_cuda_ci + register_amd_ci), so it breaks both 1-GPU nightlies:

Nightly suite	Job	Failing run	Result
Nightly Test (AMD)	nightly-test-1-gpu-unit	https://github.com/sgl-project/sglang/actions/runs/27100228380/job/80040799251	9/10 ✗
Nightly Test (Nvidia)	nightly-test-general-1-gpu-h100	https://github.com/sgl-project/sglang/actions/runs/27110545438/job/80007477055	19/20 ✗

Regression

Introduced by #26378 (commit f838adb, merged 2026-05-28), which added the gsp_group_distribution / gsp_zipf_alpha feature. That PR added the argparse options (with defaults) and the hard attribute access in from_args, but did not update get_benchmark_args in python/sglang/test/test_utils.py. Because test_gsp_multi_turn is registered nightly-only, it does not run in pre-merge PR CI, so the break only surfaced in the scheduled nightlies after the merge. cc @Jiminator

Modifications

Restore-only — not a new code change. This puts the two GSP fields back on the getattr(args, ..., default) convention that the rest of from_args (and #19363, "remove args dependency") already use. #26378 was the only place that deviated to direct attribute access; this reverts that one deviation:

-        group_distribution = args.gsp_group_distribution
-        zipf_alpha = args.gsp_zipf_alpha
+        group_distribution = getattr(args, "gsp_group_distribution", "uniform")
+        zipf_alpha = getattr(args, "gsp_zipf_alpha", None)

No behavior change for existing usage: the CLI/argparse path always supplies these attributes, so getattr returns exactly the same values, and the defaults ("uniform" / None) match both the argparse defaults and the feature's own test helper (make_args). The only difference is that an in-process Namespace that omits the fields now defaults (as the sibling fields already do) instead of raising AttributeError.

Accuracy Tests

N/A — does not affect model outputs (benchmark-client arg handling only).

Speed Tests and Profiling

N/A — pure arg-handling bugfix.

Verification

• Local repro with the exact SimpleNamespace that test_gsp_multi_turn builds via get_benchmark_args(...): before → AttributeError: ... 'gsp_group_distribution'; after → GeneratedSharedPrefixDataset(..., group_distribution='uniform', zipf_alpha=None).
• AMD CI dispatch on this branch (nightly-test-amd.yml, nightly-test-1-gpu-unit): https://github.com/sgl-project/sglang/actions/runs/27150321984/job/80139313149 → 10/10 passed (test_bench_serving_functionality.py now passes; previously 9/10).

Checklist

• Format your code according to the Format code with pre-commit.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #27149657199
Latest PR Test (Extra): 🚫 Run #27153205907

[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!

[michaelzhang-ai]

Heads-up @Jiminator: this fixes a regression from #26378. That PR added gsp_group_distribution / gsp_zipf_alpha and reads them in GeneratedSharedPrefixDataset.from_args via direct attribute access, but get_benchmark_args in python/sglang/test/test_utils.py was not updated to provide them. As a result the nightly-only test_gsp_multi_turn has been failing on both the AMD and NVIDIA 1-GPU nightlies since 2026-05-28 (it doesn't run in pre-merge PR CI).

Fix here switches the two reads to getattr(..., default) matching the argparse defaults and the other gsp_* fields. Verified on an AMD nightly dispatch (10/10). Could you take a look?

[Jiminator]

/tag-and-rerun-ci

[Jiminator]

/rerun-test test/registered/bench_fn/test_bench_serving_functionality.py::TestBenchServingFunctionality.test_gsp_multi_turn

[github-actions]

Results for /rerun-test test/registered/bench_fn/test_bench_serving_functionality.py::TestBenchServingFunctionality.test_gsp_multi_turn:

🚀 1-gpu-h100 (1 test): ✅ View workflow run

cd test/ && python3 registered/bench_fn/test_bench_serving_functionality.py TestBenchServingFunctionality.test_gsp_multi_turn

[Jiminator]

LGTM, thank you for catching this @michaelzhang-ai

[alisonshao]

closing other pr #27623, same fix