[Asthenia0412]

Summary

Fix #28808.

TRTLLMHAAttnBackend(FlashInferAttnBackend) overrode every attention-specific
method (init_forward_metadata, forward_decode, forward_extend, get_cuda_graph_seq_len_fill_value, ...).
The only real reuse from the FlashInfer parent was ~10 lines of init plumbing:
token/KV pool construction, kv_indptr/kv_last_page_len buffer allocation,
skip_prefill flag, and SWA resolution.

Inheriting from FlashInferAttnBackend additionally allocated:

• prefill_wrapper_ragged, prefill_wrappers_paged, prefill_wrappers_verify
• decode_wrappers
• FlashInferIndicesUpdater* objects
• shared global_workspace_buffer

None of these are used on the TRTLLM path. Several fields were also set twice:
_swa_kv_pool and max_context_len.

Change: TRTLLMHAAttnBackend now inherits directly from AttentionBackend.
The small amount of genuinely-shared setup is inlined. No behavioral change
intended — all public API is preserved.

Diff: 1 file, +40 / -9.

Test Plan

• Code inspection — identical public API surface, all method signatures unchanged
• python -m py_compile python/sglang/srt/layers/attention/trtllm_mha_backend.py succeeds
• No new runtime paths added

/cc @merrymercy

---

DCO: Signed-off-by: yancy yancy@users.noreply.github.com

---

CI States

Latest PR Test (Base): ❌ Run #38412114019
Latest PR Test (Extra): ❌ Run #38412113292
Latest PR Test (AMD ROCm 10): ❌ Run #38412113161

[Asthenia0412]

@merrymercy @Fridge003 @ispobock @Qiaolin-Yu @hebiao064 @HaiShaw — Hey team! Per #28808, TRTLLMHAAttnBackend was inheriting from FlashInferAttnBackend but overriding every attention method and allocating a bunch of prefill/decode wrappers plus indices updaters that the TRTLLM path never touches. This PR moves it to inherit from AttentionBackend directly — 40/9 diff, no behavior change, same public API. Take a look when you get a chance!

[Asthenia0412]

Same — needs 'run-ci' label to trigger CI tests.

[Asthenia0412]

🤖 CI gating notice: This PR's checks (call-gate/pr-gate, pr-test-*-finish) fail because this repo requires a run-ci label to trigger the actual test matrix. The failed jobs are just CI gating placeholders — real tests haven't run yet.

[hnyls2002]

Thanks for your interest in SGLang. This account opened 16 PRs in about eight hours, across unrelated subsystems (attention backends, speculative decoding, diffusion loading, scheduler IPC, KV-cache compression, docs), including two PRs for the same change (#44073 and #44117, both for #28808) and a fix for #43929 that is already on main. We can't review contributions at this volume, so we are closing the ones still open. If you want to pursue one, please pick a single change, verify it on a real run (a repro, a test, or numbers), and open a fresh PR.