GitLab

去掉DTBmm的一个冗余条件

See merge request dcutoolkit/deeplearing/vllm!416

perf: mla后面的DTbmm融合

See merge request dcutoolkit/deeplearing/vllm!415

feat:w4a8Linear调用apply_int8_linear，以支持blaslt

See merge request dcutoolkit/deeplearing/vllm!413

V0.9.2 dev du connector

See merge request dcutoolkit/deeplearing/vllm!409

fix: 修复重复判断逻辑

See merge request dcutoolkit/deeplearing/vllm!406

[P/D] add new connector as the self-developed du_swift_connector

See merge request dcutoolkit/deeplearing/vllm!405

# Conflicts:
#   vllm/distributed/kv_transfer/kv_connector/v1/p2p/p2p_nccl_engine.py

feat: Support shared experts fusion.

See merge request dcutoolkit/deeplearing/vllm!404

feat: support moe sum when topk==9

bugfix: Fix mtp model load error when eable shared experts fusion.

fix: 修复pp资源抢占bug

See merge request dcutoolkit/deeplearing/vllm!402

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