GitLab

[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

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

修改schedule_default和schedule_split_pd的选择

# 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

This reverts merge request !219

Signed-off-by: Jee Jee Li <pandaleefree@gmail.com>

[Misc] Fix `Unable to detect current VLLM config. Defaulting to NHD kv cache layout` warning (#20400)
Signed-off-by: NickLucche <nlucches@redhat.com>

Signed-off-by: Andy Xie <andy.xning@gmail.com>

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

Signed-off-by: NickLucche <nlucches@redhat.com>

Signed-off-by: mgoin <mgoin64@gmail.com>

Signed-off-by: Abatom <abzhonghua@gmail.com>

Signed-off-by: Woosuk Kwon <woosuk.kwon@berkeley.edu>

[Refactor] Create a function util and cache the results for `has_deepgemm`, `has_deepep`, `has_pplx` (#20187)
Signed-off-by: yewentao256 <zhyanwentao@126.com>

Signed-off-by: ilmarkov <imarkov@redhat.com>
Signed-off-by: Haoyang Li <Haoyang.Li@amd.com>
Co-authored-by: ilmarkov <imarkov@redhat.com>