GitLab

merge v0.11.0-nccl into v0.11.0-dev

See merge request dcutoolkit/deeplearing/vllm!305

解决gptq的不能开启graph的问题

See merge request dcutoolkit/deeplearing/vllm!304

修复CompressedTensorsLinearMethod中的w4a16的冲突问题

See merge request dcutoolkit/deeplearing/vllm!302

feat: add Marlin W16A16 fused MoE behind VLLM_USE_MARLIN_W16A16_MOE

See merge request dcutoolkit/deeplearing/vllm!300

新增环境变量 VLLM_USE_MARLIN_W16A16_MOE，用于显式启用 Marlin W16A16 MoE experts
在 fused_moe 中当开关开启且实现可用时，调用 fused_experts_impl_w16a16_marlin
增加 Marlin W16A16 MoE 实现与 reduce 路径

replace the fp8_mqa_logits and fp8_paged_mqa_logits interfaces in deepgemm with mqa_logits and paged_mqa_logits from lightop

去掉无效代码

See merge request dcutoolkit/deeplearing/vllm!285

This reverts commit 3b121add.

Add new benchmark configurations for gfx936_80cu with E=512,N=64 and E=512,N=128 Qwen3-Next-80B-A3B-Instruct nn tp4 tp8 moe json

See merge request dcutoolkit/deeplearing/vllm!283

Add new benchmark configurations for gfx936_80cu with E=512,N=64 and E=512,N=128 Qwen3-Next-80B-A3B-Instruct nn tp4 tp8 moe json

add VLLM_USE_OPT_RESHAPE_AND_CACHE、VLLM_USE_FUSE_SILU_AND_MUL and VLLM_USE_TOPK_RENORM for qwen3-30b

[fix]解决deepseek模型cudagraph模式下精度异常问题

See merge request dcutoolkit/deeplearing/vllm!278

修复schedule_default，schedule_split_pd，因kv cache耗尽，服务挂掉的问题

See merge request dcutoolkit/deeplearing/vllm!266

update VLLM_USE_PD_SPLIT=0 (for dspk)and  VLLM_USE_PD_SPLIT=1 (for others)

add VLLM_USE_PP_SYNC to use pp sync
update qwen3 of rmsnorm

[feat]支持prefill和decoding分开调度

See merge request dcutoolkit/deeplearing/vllm!260