GitLab

[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。

[feat]w4a8和w8a8适配deepep低延迟

See merge request dcutoolkit/deeplearing/vllm!255

[fix]修复mtp中的笔误

See merge request dcutoolkit/deeplearing/vllm!250

[fix]解决开启mtp后，在极端情况碰到显存不足时，导致mla中申请的tensor数据错乱问题

See merge request dcutoolkit/deeplearing/vllm!246

[fix]解决同时开启chunked-prefill和并行解码出现的指标统计错误问题

See merge request dcutoolkit/deeplearing/vllm!243

[fix]修复开启mtp并且显存不足时发生的超出维度限制问题

See merge request dcutoolkit/deeplearing/vllm!242

整合mori和deepep相关代码

See merge request dcutoolkit/deeplearing/vllm!239

# Conflicts:
#	vllm/model_executor/models/deepseek_v2.py

# 第一个ip为D的第一个节点，第二个ip为D的第二个节点,配置:export IP_CONFIG_FILE=/data/xiabo/w4a8_1/ip_config.txt
192.168.1.1 192.168.1.100
192.168.1.2 192.168.1.101
192.168.1.3 192.168.1.102
10.16.1.75 10.16.1.76