GitLab

修复schedule_default，schedule_split_pd，因kv cache耗尽，服务挂掉的问题

See merge request dcutoolkit/deeplearing/vllm!266

update VLLM_USE_PD_SPLIT=0 (for dspk)and  VLLM_USE_PD_SPLIT=1 (for others)

add VLLM_USE_PP_SYNC to use pp sync
update qwen3 of rmsnorm

[feat]支持prefill和decoding分开调度

See merge request dcutoolkit/deeplearing/vllm!260

Add minimax_m2

See merge request dcutoolkit/deeplearing/vllm!258

[fix]解决moe_fused_gate编译错误

See merge request dcutoolkit/deeplearing/vllm!254

set default_max_num_batched_tokens = 10240
update qwen3_moe of layernorm
off lightop of moe_fused_gate

add VLLM_USE_PD_SPLIT to split prefill and decode
replace triton_ of rms and act_and_mul

[fix]解决moe_fused_gate编译错误，去掉mla中mtp部分的修改

See merge request dcutoolkit/deeplearing/vllm!251

add deps of pd-pp

# 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