GitLab

fuse_moe_fp8接入marlin算子

See merge request dcutoolkit/deeplearing/vllm!399

fix(v1)：修复抢占恢复时 BlockTable 溢出

See merge request dcutoolkit/deeplearing/vllm!398

  - 缺失请求统一延后重新加入，避免同一轮重复写入
  - 抢占恢复用覆盖写入（add_row），正常运行用增量追加（append_row）
  - 保持正常请求的追加语义不变

feat(moe)：新增 VLLM_USE_MOE_W16A16_TRTION 强制 Triton MoE

See merge request dcutoolkit/deeplearing/vllm!396

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）

fix: 解决原版0消耗chunk-prefill崩溃问题

See merge request dcutoolkit/deeplearing/vllm!393

fp8增加fused_moe_gate参数

See merge request dcutoolkit/deeplearing/vllm!391

V0.9.2 dev channel lxh

See merge request dcutoolkit/deeplearing/vllm!390

[feat]添加dp attention功能

See merge request dcutoolkit/deeplearing/vllm!383

feat(moe/marlin): Marlin W16A16 MoE 自动探测并预打包（去掉手动开关）

See merge request dcutoolkit/deeplearing/vllm!382

for qwen3, VLLM_USE_FUSED_RMS_ROPE=1 (default)

  - 移除 VLLM_USE_MARLIN_W16A16_MOE 环境变量
  - 初始化阶段基于 lightop 探测并缓存 _marlin_w16a16_moe_enabled，满足条件强制 use_nn_moe=False
  - 权重加载后按缓存结果一次性 Marlin pack；运行时按 packed 标记走 Marlin fast path

perf(qwen3): 融合 q/k RMSNorm + RoPE

See merge request dcutoolkit/deeplearing/vllm!381

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

fused_moe_fp8接入lmslim

See merge request dcutoolkit/deeplearing/vllm!379

feat: Support w8a8-fp8 GEMM backend.

See merge request dcutoolkit/deeplearing/vllm!375

[qwen3-480b] MoE(TN) configs for nmz TP=8