Skip to content

GitLab

Switch branch/tag
Find file BlameHistoryPermalink
  • zhuwenwen's avatar
    perf(qwen3): 融合 q/k RMSNorm + RoPE · 3cd3d1e6
    zhuwenwen authored 
    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
    3cd3d1e6
qwen3.py 15.9 KB