GitLab

[feat]优化dp attention，减少1次allgather耗时，高吞吐提升明显

See merge request dcutoolkit/deeplearing/vllm!307

w8a8 高吞吐模式先量化再dispatch

See merge request dcutoolkit/deeplearing/vllm!303

[feat]修复低延迟错误

See merge request dcutoolkit/deeplearing/vllm!301

优化高吞吐模式num_sms

See merge request dcutoolkit/deeplearing/vllm!297

[feat]1.支持高吞吐模式ep_scatter+deepgemm contiguous+ep_gather方案；2.支持高吞吐模式下ETP,例如dp4 tp4

See merge request dcutoolkit/deeplearing/vllm!296

[fix]修复deepep 高吞吐模式vmfault问题

See merge request dcutoolkit/deeplearing/vllm!291

[feat]支持deepep ETP，dp4 tp4 ep16相比dp32 tp1 ep32提升明显

See merge request dcutoolkit/deeplearing/vllm!289

add ALLOW_MNNV default falase, use VLLM_ALLOW_MNNVL=1

See merge request dcutoolkit/deeplearing/vllm!286

[feat]支持deepep低延迟与共享专家overlap

See merge request dcutoolkit/deeplearing/vllm!281

feat: pp mtp加入零消耗调度，加入环境变量VLLM_USE_ZERO_MTP，默认打开

See merge request dcutoolkit/deeplearing/vllm!277

feat: pipeline_parallel新增pp域请求数均衡,VLLM_USE_PP_BALANCE控制，默认开启

See merge request dcutoolkit/deeplearing/vllm!267

[feat]1.w8a8 marlin适配deepep低延迟;2.非naive ep模式，去掉多余的dp padding,避免allreduce耗时

See merge request dcutoolkit/deeplearing/vllm!256

restore the default settings of disable_cascade_attn
add VLLM_USE_OPT_ZEROS to replace triton_ (torch.zeros)
set default_max_num_batched_tokens = 10240
update qwen3_moe of layernorm

解决w8a8 pp16开启marlin的oom问题

- 在环境变量中引入 `VLLM_SCHED_ENABLE_MINIMAL_INJECTION` 以控制流水线并行调度的最小注入。
- 调整 Scheduler 逻辑以使用新的最小注入功能。
- 更新调度逻辑以利用输出占位符，确保在解码过程中避免 0-token 停滞。
- 增强 Scheduler，根据批次队列状态管理最小进度注入。