GitLab

  - 移除 VLLM_USE_MARLIN_W16A16_MOE 环境变量
  - 初始化阶段基于 lightop 探测并缓存 _marlin_w16a16_moe_enabled，满足条件强制 use_nn_moe=False
  - 权重加载后按缓存结果一次性 Marlin pack；运行时按 packed 标记走 Marlin fast path

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

feat: Support w8a8-fp8 GEMM backend.

See merge request dcutoolkit/deeplearing/vllm!375

[qwen3-480b] MoE(TN) configs for nmz TP=8

add VLLM_USE_FUSED_CACHE_QUANT_BMM_MLA to use fused rmsnorm + contiguous + rope(for dpsk-v3) + concat_and_cache_mla + q quant, control bmm(todo) + cat +mla (fp8)

新增环境变量 VLLM_MOE_ROUTER_CAPTURE / DIR / RANK / MAX_LAYERS / NUM_TOKENS_* 用于开关与过滤控制
新增 router_capture.py，支持按 num_tokens 分桶抓取 router logits 并落盘
在 qwen3_moe 中接入抓取逻辑，默认关闭，仅在开启时记录
固定 skip_profile / skip_stack_funcs 为默认启用，避免抓到 warmup/profile 形状
统一配置入口到 vllm.envs，作为运行时基准

Switch default w8a8 gemm impl to blaslt.

See merge request dcutoolkit/deeplearing/vllm!368

V0.9.2 dev yql 1.15

See merge request dcutoolkit/deeplearing/vllm!367

vllm：export VLLM_CUSTOM_CACHE=1
dtk：export HIP_KERNEL_EVENT_SYSTENFENCE=1

V0.9.2 dev fix pp+cp

See merge request dcutoolkit/deeplearing/vllm!364

fix PP 场景 decode 阶段 token 被误丢弃导致卡住

See merge request dcutoolkit/deeplearing/vllm!363

  - decode 已开始时不再按 partial prefill 丢弃 sampled token，避免 new_token_ids=[] 循环拖尾

修改非堆成切分的判断

See merge request dcutoolkit/deeplearing/vllm!362

nmz适配block和channel fp8

See merge request dcutoolkit/deeplearing/vllm!360