GitLab

  参考并移植 011/vllm 的关键提交逻辑
  新增 VLLM_USE_MOE_W16A16_TRITON 开关，并接入基于 lightop 的运行时能力探测与启用结果缓存。
  在权重加载后对 w13 与 w2 执行 W16A16 Marlin 预打包。
  W16A16 Marlin 启用时保留 monolithic 执行路径，并在 fused_experts_impl 中增加 packed 权重 fast-path。
  保持 Marlin 或 lightop 不可用时的回退行为不变。

[feat]宽松mtp支持temp,top-p等参数设置

See merge request dcutoolkit/deeplearing/vllm!420

[feat]支持prefill和decode调度分离

See merge request dcutoolkit/deeplearing/vllm!419

适配w8a8 deepep,接入lightop版deepgemm

See merge request dcutoolkit/deeplearing/vllm!418

set fp8_e4m3 only supported on nmz and support q&kvcache fp8
set VLLM_PCIE_USE_CUSTOM_ALLREDUCE=1

[feat]支持宽松mtp

See merge request dcutoolkit/deeplearing/vllm!414

# Conflicts:
#	vllm/model_executor/layers/fused_moe/modular_kernel.py

# Conflicts:
#	vllm/model_executor/layers/fused_moe/config.py
#	vllm/model_executor/layers/fused_moe/layer.py
#	vllm/model_executor/layers/quantization/compressed_tensors/compressed_tensors_marlin.py