GitLab

[feat] 支持 mRoPE 的 fused RMSNorm+RoPE 路径，并修正 torch.compile 动态维度标注

See merge request dcutoolkit/deeplearing/vllm!334

实现了用于优化张量计算的 rms_mrope_fuse 和 rms_mrope_fuse_fake 方法
更新了 forward：在满足条件时走新的 M-RoPE 融合路径
增强了 Qwen3MoeModel 对动态参数维度的支持，以适配该功能

在v0.11中暂不支持w4a16的moe算子scale和zero合并操作

See merge request dcutoolkit/deeplearing/vllm!318

[PD][Feat]支持fa_pa kvcahe类型模型推理

See merge request dcutoolkit/deeplearing/vllm!317

set VLLM_USE_PIECEWISE=0

适配gptq/awq的triton moe算子

See merge request dcutoolkit/deeplearing/vllm!313

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