GitLab

移除 forward 中对 experts.use_overlapped/_shared_experts 的状态改写，避免 torch.compile 启动期 shared/non-shared 路径不一致
FusedMoE.forward_impl 仅在 shared_output 为空时计算 shared experts，防止透传值被本地重算覆盖

  新增环境变量 VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD 用于控制
  fused sum+mul+add 开关。
  在 DeepseekV2MoE 中增加 fused 路径，预计算 shared_output，并下传 iqis 与 routed_scaling_factor。
  扩展 FusedMoE/SharedFusedMoE 及相关 custom op 接口，统一透传 i_q/i_s/shared_output/routed_scaling_factor。
  同步适配 Triton、Marlin W16A16、SlimQuant W4A8、CompressedTensors W8A8 等实现，支持在内核侧完成 sum+mul+add。

add fa unified attn 导入判断

See merge request dcutoolkit/deeplearing/vllm!509

invoke flash_attn in the Qwen2AudioEncoder (transformers)

See merge request dcutoolkit/deeplearing/vllm!508

fix: resolve block_shape conflicts between DeepEP MoE and non-DeepEP quantization

See merge request dcutoolkit/deeplearing/vllm!507

Merge v0.15.1-dev-pd into v0.15.1-dev

See merge request dcutoolkit/deeplearing/vllm!506

Fix Qwen3/Qwen3.5 Reasoning Parser  (#34779)

See merge request dcutoolkit/deeplearing/vllm!504

perf: GLM4.7增加MOE调用rmsQuant, fix: 修掉fused_moe向后传递None导致的报错

See merge request dcutoolkit/deeplearing/vllm!505

Add FA Unified Attention 2D

See merge request dcutoolkit/deeplearing/vllm!501

[feat]deepseek mtp支持pp模式

See merge request dcutoolkit/deeplearing/vllm!503

[feat]支持ray分布式异步调度，VLLM_ENABLE_RAY_ASYNC_SCHEDULING环境变量控制

See merge request dcutoolkit/deeplearing/vllm!502

- Add VLLM_V1_USE_FA_UNIFIED_ATTN_2D 环境变量
- 0: Triton attention, 1: FA unified attention

rms_norm_opt精度问题解决（换了个kernel）

See merge request dcutoolkit/deeplearing/vllm!499

fix: 修复MOE量化tensor对于其他模型的影响

See merge request dcutoolkit/deeplearing/vllm!500

修改sparse_attn hip后端

See merge request dcutoolkit/deeplearing/vllm!498

[fix]添加VLLM_USE_LIGHTOP_FUSED_TOPP_TOPK控制lightop topp_topk融合算子开关

See merge request dcutoolkit/deeplearing/vllm!496

Fix：GLM-5量化模型mla_attention layout修复&&sparse_attn fp8支持

See merge request dcutoolkit/deeplearing/vllm!495

feat(deepseek-mla): 接入 VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT 融合链路

See merge request dcutoolkit/deeplearing/vllm!486

新增环境变量与 MLA 融合接线（wrapper -> attention -> impl） 接入 lightop fused_rms_norm_rope_contiguous，保留回退路径

moe: 补齐 fill+moe_align 融合开关语义

See merge request dcutoolkit/deeplearing/vllm!484