GitLab

支持kvacache fp8_e4m3的RMS_ROPE_CONCAT

See merge request dcutoolkit/deeplearing/vllm!531

处理VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD导致的awq推理bug问题

See merge request dcutoolkit/deeplearing/vllm!528

fix(moe): 补齐非Marlin量化路径 shared_output/routed_scaling_factor 透传

See merge request dcutoolkit/deeplearing/vllm!529

支持kvacache fp8_e4m3/fp8_e5m2的RMS_ROPE_CONCAT

Gy v015 1dmrope

See merge request dcutoolkit/deeplearing/vllm!530

支持kvacache fp8_e4m3/fp8_e5m2的RMS_ROPE_CONCAT

删除1d_mrope

See merge request dcutoolkit/deeplearing/vllm!525

关闭sparse_mla的num_head到64/128的pad，以及添加控制fp8_use_mixed_batch模式的环境变量控制，FP8_USE_MI...

See merge request dcutoolkit/deeplearing/vllm!524

# Conflicts:
#   vllm/model_executor/layers/sparse_attn_indexer.py
#   vllm/v1/attention/backends/mla/flashmla_sparse.py

feat:flash_mla，q去掉pad

See merge request dcutoolkit/deeplearing/vllm!522

关闭sparse_mla的num_head到64/128的pad，以及添加控制fp8_use_mixed_batch模式的环境变量控制，FP8_USE_MIXED_BATCH，默认为false，为分离模式

[perf]DSA架构模型支持mtp>1

See merge request dcutoolkit/deeplearing/vllm!521

fix(moe): 仅在 fused moe_sum+mul+add 开启时透传 shared_output

See merge request dcutoolkit/deeplearing/vllm!520

[fix]修复GLM mtp精度问题

See merge request dcutoolkit/deeplearing/vllm!518

qwen3.py合入fused_morpe

See merge request dcutoolkit/deeplearing/vllm!516

feat(moe): 修复 shared_output 透传被覆盖并兼容 torch.compile 启动路径

See merge request dcutoolkit/deeplearing/vllm!517

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

x接入mla_cat算子仅在nmz和kvcache-fp8情况下生效，默认关闭，开启需要export VLLM_USE_CAT_MLA=1

See merge request dcutoolkit/deeplearing/vllm!513

fix prompt_is_reasoning_end_arr not defined

See merge request dcutoolkit/deeplearing/vllm!515

feat(deepseek-moe): 接入 VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD 融合链路

See merge request dcutoolkit/deeplearing/vllm!485

  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。

feat:支持mqa的fp8实现

See merge request dcutoolkit/deeplearing/vllm!514