[Asthenia0412]

Problem

TRTLLMHAAttnBackend 继承自 FlashInferAttnBackend，但实际上只复用了少量 init 代码。这导致每次实例化 TRTLLMHAAttnBackend 时都会浪费大量资源：

• 分配 FlashInfer prefill_wrapper_ragged, prefill_wrappers_paged, prefill_wrappers_verify, decode_wrappers (BatchPrefillWithPagedKVCacheWrapper, BatchDecodeWithPagedKVCacheWrapper 等)
• 创建 FlashInferIndicesUpdaterPrefill / FlashInferIndicesUpdaterDecode
• 分配 kv_indptr, kv_last_page_len, qo_indptr 等 FlashInfer 专用缓冲区
• 设置 decode_use_tensor_cores, num_wrappers, dispatch_reason, enable_deterministic 等 TRTLLM 从不使用的标志

TRTLLMHAAttnMultiStepDraftBackend 继承自 FlashInferMultiStepDraftBackend，后者同样不是 AttentionBackend 子类，只是实现了 init_forward_metadata/init_cuda_graph_state 等接口方法。

Solution

TRTLLMHAAttnBackend

• 改为直接继承 AttentionBackend (base class)
• 将真正共享的池/翻译器/量化器初始化代码直接内联到 init 中
• 将 _kv_write_scales 方法 (仅 3 行) 复制到本类
• 在自身 _swa_kv_pool 赋值后直接设置 use_sliding_window_kv_pool
• 删除了 flashinfer_backend.py 的 import 依赖

TRTLLMHAAttnMultiStepDraftBackend

• 独立化 (不再继承 FlashInferMultiStepDraftBackend)
• 内联了 init 中 kv_indptr/kv_last_page_len 的分配逻辑
• 不再加载 generate_draft_decode_kv_indices, common_template 等 FlashInfer 专用逻辑

Changes

仅修改了一个文件: python/sglang/srt/layers/attention/trtllm_mha_backend.py (+52 / -18)

Testing

• Python 语法检查通过 (ast.parse)
• 所有外部调用者 (attention_registry, draft_utils, eagle_worker_v2 等) 的导入路径不变, 行为不变

Fixes: #28808
Author: yancy (Asthenia0412)

---

CI States

Latest PR Test (Base): ❌ Run #38461119613
Latest PR Test (Extra): ❌ Run #38461119568
Latest PR Test (AMD ROCm 10): ❌ Run #38461119645

[Asthenia0412]

Hey @reviewers! I've refactored TRTLLMHAAttnBackend and TRTLLMHAAttnMultiStepDraftBackend to inherit directly from AttentionBackend instead of FlashInferAttnBackend. The FlashInfer inheritance was a massive waste since TRTLLM overrides 95% of parent methods and never uses FlashInfer wrappers/updaters. Just inherited what TRTLLM actually needs (pool setup + kv_indptr buffers).

Could you take a quick look? The change is localized to one file and fully backward-compatible.

[Asthenia0412]

Hi! The CI currently shows skipped jobs because this repo requires the run-ci label to trigger the full test suite. The failing jobs you see are just the CI gating placeholders (call-gate/pr-gate, pr-test-*-finish) — real tests haven't run yet.

Could someone please add the run-ci label to kick off the actual test matrix? Thanks so much! 🙏

@lm-sys @zhyncs @hnyls2002 @ispobock @BBuf @Fridge003 @kevin-mii

[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.