[hnyls2002]

Summary

• Follow up on [Bug] Include DSA variant in exact-bucket graph admission #38992.
• Move DSA graph enumeration and batch-length selection into the attention layer; the decode runner consumes ordered capture labels and a selector.
• Generalize the independent ShapeKey.attention_variant axis and capture flag, preserving LoRA composition, draft-runner defaults, capture order, and DSA platform/fallback rules.
• Split the shared runtime foundation from [Refactor] Separate V4.1 indexer capture options from candidate state #38962 so dsv4.1: remaining model and runtime integration #38798 can later add V4.1 variants without model-specific branches in the decode runner.

---

CI States

Latest PR Test (Base): 🚫 Run #34564089878
Latest PR Test (Extra): ❌ Run #34564089745
Latest PR Test (AMD ROCm 10): ❌ Run #34564089828

[hnyls2002]

/rerun-test test/registered/unit/model_executor/runner/test_decode_cuda_graph_runner.py test/registered/unit/model_executor/runner_backend/test_full_cuda_graph_backend.py test/registered/unit/spec/test_eagle_draft_cuda_graph_runner.py test/registered/unit/model_executor/test_prefill_cuda_graph_runner_helpers.py

[github-actions]

Results for /rerun-test test/registered/unit/model_executor/runner/test_decode_cuda_graph_runner.py test/registered/unit/model_executor/runner_backend/test_full_cuda_graph_backend.py test/registered/unit/spec/test_eagle_draft_cuda_graph_runner.py test/registered/unit/model_executor/test_prefill_cuda_graph_runner_helpers.py:

🚀 ubuntu-latest (4 tests): ✅ View workflow run

cd test/ && python3 registered/unit/model_executor/runner/test_decode_cuda_graph_runner.py
cd test/ && python3 registered/unit/model_executor/runner_backend/test_full_cuda_graph_backend.py
cd test/ && python3 registered/unit/spec/test_eagle_draft_cuda_graph_runner.py
cd test/ && python3 registered/unit/model_executor/test_prefill_cuda_graph_runner_helpers.py