GitLab

  - 在 vLLM _C 扩展中新增 rms_rotary_embedding_fuse（注册 op + CUDA kernel），减少对 LightOp 的硬依赖
  - 新增环境变量 VLLM_FUSED_RMS_ROPE_BACKEND=auto|vllm|lightop，auto 优先走 vLLM，缺失时回退 LightOp
  - 更新 Qwen3 / Qwen3-MoE 的 fused 路径按后端选择执行
  - 补充 tc_opt benchmark 结果解析脚本 benchmarks/tc_opt/test/parse_bench_results.py

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）

for qwen3, VLLM_USE_FUSED_RMS_ROPE=1 (default)

  - 使用 get_moe_cuda_marlin_config_w16a16(status) 判断 W16A16 Marlin MoE 是否可用
  - 在 FusedMoE 初始化阶段计算并缓存 _marlin_w16a16_moe_enabled，满足条件时强制 use_nn_moe=False
  - 权重加载后按缓存结果进行一次性 Marlin pack；运行时按 packed 标记走 Marlin fast path
  - 删除 envs.py 中 VLLM_USE_MARLIN_W16A16_MOE 环境变量定义与解析逻辑

   vllm：export VLLM_CUSTOM_CACHE=1
   dtk：export HIP_KERNEL_EVENT_SYSTENFENCE=1
2、kvcache支持fp8

todo: add VLLM_USE_QUERY_QUANT to not use q quant

set VLLM_USE_PD_SPLIT=1
update moe_align_block_size

区分pcie和hglink custom allreduce的使用
vllm：export VLLM_CUSTOM_CACHE=1
dtk：export HIP_KERNEL_EVENT_SYSTENFENCE=1

 - fused_experts_impl 增加 VLLM_USE_MARLIN_W16A16_MOE fast path：首次对 w1/w2 做 Marlin pack 后缓存，避免重复 reorder；并将原始
    权重 offload 到 CPU，降低 GPU 双份驻留
  - envs 补齐环境变量 VLLM_USE_LIGHTOP_MOE_SUM_MUL_ADD
  - 更新 fuse_moe_w16a16_marlin.py 的报错提示为 VLLM_USE_LIGHTOP=1

新增环境变量开关 VLLM_V1_FAST_TOKEN_ID_COPY（默认关闭）。开启后在 CachedRequestState 中缓存 prompt_token_ids 的 np.int32，并在 add_request 里用 np.copyto 写入 token_ids_cpu，避免长 prompt 场景反复 list->NumPy 转换开销（尤其是抢占/反复进出 batch 时）

set VLLM_USE_PIECEWISE=0

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

add VLLM_USE_OPT_RESHAPE_AND_CACHE、VLLM_USE_FUSE_SILU_AND_MUL and VLLM_USE_TOPK_RENORM for qwen3-30b

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

set default_max_num_batched_tokens = 10240
update qwen3_moe of layernorm
off lightop of moe_fused_gate

add VLLM_USE_PD_SPLIT to split prefill and decode
replace triton_ of rms and act_and_mul

Signed-off-by: Chenheli Hua <huachenheli@outlook.com>
Signed-off-by: simon-mo <simon.mo@hey.com>

Signed-off-by: Aleksandr Malyshev <maleksan@amd.com>
Co-authored-by: Aleksandr Malyshev <maleksan@amd.com>
Co-authored-by: Doug Lehr <douglehr@amd.com>

Signed-off-by: Saman Keon <samanamp@outlook.com>

Signed-off-by: Duncan Moss <djm.moss@gmail.com>

Signed-off-by: Shu Wang <shuw@nvidia.com>
Signed-off-by: Shu Wang. <shuw@nvidia.com>
Signed-off-by: Tyler Michael Smith <tyler@neuralmagic.com>
Signed-off-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>
Co-authored-by: Tyler Michael Smith <tyler@neuralmagic.com>
Co-authored-by: Tyler Michael Smith <tlrmchlsmth@gmail.com>

Signed-off-by: rongfu.leng <rongfu.leng@daocloud.io>

Signed-off-by: Lucas Wilkinson <lwilkins@redhat.com>

Signed-off-by: Corey Lowman <clowman1993@gmail.com>