GitLab

  - 在 vLLM _C 扩展中新增 rms_rotary_embedding_fuse（注册 op + CUDA kernel），减少对 LightOp 的硬依赖
  - 新增环境变量 VLLM_FUSED_RMS_ROPE_BACKEND=auto|vllm|lightop，auto 优先走 vLLM，缺失时回退 LightOp
  - 更新 Qwen3 / Qwen3-MoE 的 fused 路径按后端选择执行
  - 补充 tc_opt benchmark 结果解析脚本 benchmarks/tc_opt/test/parse_bench_results.py

V0.11.0 dev lxh channelwise

See merge request dcutoolkit/deeplearing/vllm!401

feat(moe)：新增 VLLM_USE_MOE_W16A16_TRTION 强制 Triton MoE

See merge request dcutoolkit/deeplearing/vllm!397

增加环境变量开关，禁用 Marlin W16A16 MoE 路径
强制 Triton 且权重已是 Marlin packed 时给出明确报错
Marlin 支持探测改为 best-effort（不再依赖 VLLM_USE_LIGHTOP）

V0.11.0 dev q

See merge request dcutoolkit/deeplearing/vllm!392

fused_moe_fp8接入lmslim

See merge request dcutoolkit/deeplearing/vllm!389

[MiniMax-M2.1] 新增minimax-m2.1工具调用

See merge request dcutoolkit/deeplearing/vllm!387

[fix]解决dp+graph+naive all2all dispatch时的维度不匹配问题

See merge request dcutoolkit/deeplearing/vllm!384

新增 VLLM_USE_FUSED_RMS_ROPE 分支，走 fused 路径
注册 torch.ops.vllm.rms_rotary_embedding_fuse（direct_register_custom_op）
cos_sin_cache 自动转 device/dtype 并缓存，避免每次重复拷贝

for qwen3, VLLM_USE_FUSED_RMS_ROPE=1 (default)

feat: Supprot fp8 channle-wise matmul.

See merge request dcutoolkit/deeplearing/vllm!380

V0.11.0 dev kvscale

See merge request dcutoolkit/deeplearing/vllm!378

feat: support fp8-blockwise matmul impl.

See merge request dcutoolkit/deeplearing/vllm!377

feat(moe/marlin): 移除 VLLM_USE_MARLIN_W16A16_MOE，改为基于 lightop 探测自动启用并一次性缓存决策

See merge request dcutoolkit/deeplearing/vllm!376

  - 使用 get_moe_cuda_marlin_config_w16a16(status) 判断 W16A16 Marlin MoE 是否可用
  - 在 FusedMoE 初始化阶段计算并缓存 _marlin_w16a16_moe_enabled，满足条件时强制 use_nn_moe=False
  - 权重加载后按缓存结果进行一次性 Marlin pack；运行时按 packed 标记走 Marlin fast path
  - 删除 envs.py 中 VLLM_USE_MARLIN_W16A16_MOE 环境变量定义与解析逻辑

[fix]解决glm4 moe + mtp精度异常

See merge request dcutoolkit/deeplearing/vllm!374