GitLab

[perf]合入lightop topp_topk 融合算子

See merge request dcutoolkit/deeplearing/vllm!456

修复awq-marlin的bug

See merge request dcutoolkit/deeplearing/vllm!454

fix(qwen3): 修复VL场景fused RoPE分支条件，并更正qwen3_moe中rms_mrope参数顺序

See merge request dcutoolkit/deeplearing/vllm!451

[perf]优化异步调度+并行解码 step之间的空泡，实现kernel提前下发

See merge request dcutoolkit/deeplearing/vllm!449

修复qwen3.5的使用dtype为fp16的picecwise的推理模式

See merge request dcutoolkit/deeplearing/vllm!448

feat(sampler): 增加 reduced topk+topp 采样快速路径以降低全词表 softmax 开销

See merge request dcutoolkit/deeplearing/vllm!447

新增 VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER 开关并补充适用场景说明

在 V1 GPU 输入批预计算 max_top_k/has_any_no_top_k，native sampler 满足条件时走 reduced fast path，异常自动回退

修复block-wise 用triton报错的bug

See merge request dcutoolkit/deeplearing/vllm!444

修复channel-wise 用triton报错的bug

See merge request dcutoolkit/deeplearing/vllm!443

V0.15.1 dev w4a8+pp balance

See merge request dcutoolkit/deeplearing/vllm!442

# Conflicts:
#   vllm/envs.py

feat(moe): 支持通过环境变量开启/配置 Qwen3 路由 logits 采集

See merge request dcutoolkit/deeplearing/vllm!441

perf(v1): 增加可选的快速 token-id 拷贝路径

See merge request dcutoolkit/deeplearing/vllm!440

新增 router_capture 工具，用于按 num_tokens/rank 过滤并落盘 MoE router logits
在 Qwen3MoeSparseMoeBlock 中接入采集调用，并在 torch.compile 场景下自动跳过
补充 VLLM_MOE_ROUTER_CAPTURE* 环境变量

  - 新增环境变量 `VLLM_V1_FAST_TOKEN_ID_COPY`（默认关闭）
  - 在 `CachedRequestState` 中缓存 int32 的 prompt token ids（numpy 数组）
  - 开启后在 `InputBatch` 中使用 `np.copyto` 拷贝 prompt/output token ids

Support qwen3 5

See merge request dcutoolkit/deeplearing/vllm!438

支持非deepseek模型使用moe_fused_gate

See merge request dcutoolkit/deeplearing/vllm!439

[feat]支持glm4_moe_mtp使用torch compile，实现mtp cudagraph模式

See merge request dcutoolkit/deeplearing/vllm!436

feat(moe): 补齐 v0.15 中 Marlin W16A16 MoE 端到端接入

See merge request dcutoolkit/deeplearing/vllm!431