GitLab

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）

perf: DS-量化模型融合qa和kva的gemm

See merge request dcutoolkit/deeplearing/vllm!329

fix: chunk-prefill生成的token被错误返回

See merge request dcutoolkit/deeplearing/vllm!327

V0.9.2 dev main+mtp zero

See merge request dcutoolkit/deeplearing/vllm!325

V1 采样器：新增 reduced top-k/top-p 采样路径

See merge request dcutoolkit/deeplearing/vllm!321

修复w8a8 triton config 择优位运算可能引发torch compile 编译错误，修复smquant w8a8 权重后处理位置

See merge request dcutoolkit/deeplearing/vllm!320

# Conflicts:
#	vllm/envs.py

新增环境变量 VLLM_V1_USE_REDUCED_TOPK_TOPP_SAMPLER 用于开关控制
扩展 SamplingMetadata，增加 max_top_k 与 has_any_no_top_k
在 InputBatch 侧计算 top-k 的主机端汇总信息，避免 device 同步
更新 Sampler/TopKTopPSampler 传递并使用新参数以启用优化采样

mla模型P、D单实例单机的任意切分方式（满足D的tp>=P的tp)使用

See merge request dcutoolkit/deeplearing/vllm!315

update VLLM_USE_LIGHTOP_RMS_ROPE_CONCAT impl

[feat]1.优化ep sequence parallel,区分主模型和mtp逻辑;2.ep sequence parallel添加cudagraph padding到tp_size;3.修复共享专家和deepep combine overlap

支持pd分离:P做跨机pp2tp8，D做单机的tp8.

See merge request dcutoolkit/deeplearing/vllm!311

使用方式P在connector的参数添加："enable_asymmetric_p2p":true,"remote_tp_size":8,"remote_pp_size":1 这3个参数。例：
"kv_connector_extra_config":{"enable_asymmetric_p2p":true,"remote_tp_size":8,"remote_pp_size":1,"proxy_ip":"10.0.1.1","proxy_port":"30001","http_port":"20001"}