[Bugfix] HybridAttnBackend declares forward() twice, killing the linear-attention path

[noron12234]

Motivation

HybridAttnBackend declares forward() twice. Python binds the later
definition, so the earlier one is dead code:

line	origin	signature
154	[Kimi] Support kimi-k3 (#32541), 2026-08-04	q, k, v, layer, forward_batch, save_kv_cache, *, mixed_qkv, a, b, **kwargs
224	[feat] Support different attention backends for prefill and decode, 2025-07-28	q, k, v, layer, forward_batch, save_kv_cache, **kwargs

#32541 added the linear-attention-aware forward() at the top of the class but
did not remove the pre-existing one further down, so the 2025 definition is the
one that is actually bound.

The effect is that the linear-attention dispatch added by #32541 never runs.
A caller passing mixed_qkv / a / b — as
hybrid_linear_attn_backend.py does — falls into **kwargs on the old
definition and is then handed to backend.forward() alongside q=None, k=None, v=None, instead of taking the if mixed_qkv is not None: branch.

Modifications

• Remove the stale 2025 forward(). The kimi-k3 one is a strict superset: same
  parameters for the regular path, plus the linear-attention branch.
• Add two unit tests to
  test/registered/unit/model_executor/model_runner_components/test_attention_backend_setup.py
  (already a CPU test module, and it already imports HybridAttnBackend)
  asserting that forward() routes mixed_qkv/a/b to the selected backend
  and that the plain q/k/v path is unchanged. A future duplicate
  definition fails these instead of silently shadowing.

No behaviour change for the regular attention path; the only difference is
positional vs keyword argument passing to backend.forward().

Accuracy Tests

Not applicable — this restores a dispatch branch that is currently unreachable,
it does not change kernel or model math.

Verification

black and isort clean on both touched files.

I was not able to run the full test module locally: importing sglang.srt on
this machine fails at transformers config registration
(ValueError: 'qwen3_asr' is already used by a Transformers config), which is a
version-pinning issue in my environment and unrelated to this change. I
verified the dispatch by executing the real, unmodified HybridAttnBackend
class body with stubbed imports:

forward definitions in class: 1
linear path  -> prefill {'mixed_qkv': 'mixed', 'a': 'a', 'b': 'b'}
regular path -> ('q', 'k', 'v')

Please run the added tests in CI.

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.
• Follow the SGLang code style guidance.

---

CI States

Latest PR Test (Base): ❌ Run #33099494579
Latest PR Test (Extra): ❌ Run #33099494294
Latest PR Test (AMD ROCm 7.2): ❌ Run #33099494501

[noron12234]

Rebased onto current main; the branch was well over a thousand commits behind and every red check was an aggregator (pr-gate, call-gate, *-finish) from a stale run rather than a leaf test failure. Cherry-pick applied clean onto the current tip, 2 files. black --check clean on both. Fresh CI should give a real signal now.

[noron12234]

Closing this. I am cleaning up the PRs I have open and keeping only the ones a maintainer has already picked up; nobody has needed to look at this one, and leaving it open just adds to your queue.

The branch stays on my fork, so nothing is lost if the fix turns out to be worth having. Sorry for the noise.