GitLab

set VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT=1

feat: 为Qwen3 MoE添加RMSNorm和RoPE融合优化与qwen3-480B tp8 moe配置文件

See merge request dcutoolkit/deeplearing/vllm!295

- 新增rms_rotary_embedding_fuse自定义操作
- 添加内核配置文件E=160,N=320
- 通过VLLM_USE_FUSED_RMS_ROPE环境变量控制融合路径

feat: add Marlin W16A16 MoE fast path

See merge request dcutoolkit/deeplearing/vllm!294

- 添加环境变量开关 VLLM_USE_MARLIN_W16A16_MOE（及 fused RMS+RoPE 标志位）
- 增加 W16A16 Marlin 权重重排/缓存逻辑
- 集成 lightop 的 W16A16 MoE fast path，复用打包缓存避免重复 pack 并在 GPU 卸载原布局

fix: 细化量化模型开启融合场景

See merge request dcutoolkit/deeplearing/vllm!292

[feat]零消耗适配宽松mtp

See merge request dcutoolkit/deeplearing/vllm!290

set USE_FUSED_RMS_QUANT=1, USE_FUSED_SILU_MUL_QUANT=1 and  VLLM_SCHED_ENABLE_MINIMAL_INJECTION=1 for dpsk-v3

set VLLM_REJECT_SAMPLE_OPT=1 for dpsk-v3

- 在 InputBatch.refresh_metadata 中为展开后的采样元数据引入 repeat_count 记录重复次数
- 完善元数据刷新逻辑以适配 reject-sampling 优化路径
- 更新 GPUModelRunnerBase，在 batch 处理阶段正确消费新的采样元数据与重复计数

[fix]解决mtp保存draft prob在例如pd分离场景下的OOM问题

See merge request dcutoolkit/deeplearing/vllm!284

add VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT when use USE_FUSED_RMS_QUANT and USE_FUSED_CUSTOM_ALL_REDUCE_RMS_QUANT

回退错误修改，避免dp模式漏掉的dispatch和combine通信

See merge request dcutoolkit/deeplearing/vllm!282

[fix]解决dp模式mtp卡住问题

See merge request dcutoolkit/deeplearing/vllm!280