[MiniMax-M3] Fix UnboundLocalError in the sparse backend on the MSA path

[David-Wu1119]

Motivation

MiniMaxSparseAttnBackend.__init__ calls get_parallel() when use_msa is set (self.num_q_heads = ... // get_parallel().attn_tp_size). get_parallel is imported at module level. #36527 later added from sglang.srt.runtime_context import get_parallel inside the index_cache_enabled branch of the same method. That import binds the name inside __init__, so Python treats get_parallel as a local for the whole method. The earlier MSA branch then reads an unbound local:

UnboundLocalError: cannot access local variable 'get_parallel' where it is not associated with a value

use_msa is true on CUDA/ROCm whenever MSA is available and the config matches (block size 128, page size 128, top-k blocks 4/8/16/32, and so on), so a MiniMax-M3 start that takes the MSA path fails in the backend constructor. This is true whether or not the index cache is enabled, because the scoping is decided when the function is compiled. You can check it without a GPU:

from sglang.srt.layers.attention.minimax_sparse_backend import MiniMaxSparseAttnBackend
"get_parallel" in MiniMaxSparseAttnBackend.__init__.__code__.co_varnames  # main: True, this PR: False

Modifications

• Remove the redundant function-level import. The module-level get_parallel is already in scope, and the index-cache branch keeps working with it.
• Add test/registered/unit/layers/attention/test_minimax_sparse_backend_scoping.py (CPU suite), in the style of test_dsa_head_gate_guard.py. No CPU runner constructs this backend, so the test checks that get_parallel is not a local of __init__. It fails on main ('get_parallel' unexpectedly found in (...)) and passes here.

python test/registered/unit/layers/attention/test_minimax_sparse_backend_scoping.py   # main: FAILED (errors=1), this PR: OK
pre-commit run --files <both files>                                                  # isort, ruff, ruff format, codespell, CI-registry checks pass

Accuracy Tests

Not applicable. No kernel or math changes.

Speed Tests and Profiling

Not applicable.

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.

cc @zcnrex (author of #36527)

🤖 Generated with Claude Code

---

CI States

Latest PR Test (Base): ❌ Run #36269536562
Latest PR Test (Extra): ❌ Run #36269536712
Latest PR Test (AMD ROCm 10): ❌ Run #36269536465

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.