GitLab

1、pcie 解决custom cudagraph模式需要拷贝的问题，这个需要配合dtk进行使用。

See merge request dcutoolkit/deeplearing/vllm!369

   vllm：export VLLM_CUSTOM_CACHE=1
   dtk：export HIP_KERNEL_EVENT_SYSTENFENCE=1
2、kvcache支持fp8

todo: add VLLM_USE_QUERY_QUANT to not use q quant

适配block-wise fp8接口

See merge request dcutoolkit/deeplearing/vllm!366

set VLLM_USE_PD_SPLIT=1
update moe_align_block_size

Switch default w8a8 gemm impl to blaslt.

See merge request dcutoolkit/deeplearing/vllm!365

处理test_concat的ds_cat的导入bug

See merge request dcutoolkit/deeplearing/vllm!361

perf(fused-moe): 预打包 Marlin W16A16 MoE 权重，降低 warmup 显存峰值

See merge request dcutoolkit/deeplearing/vllm!358

在 post-load hook 中对 w13/w2 做 per-expert Marlin pack，并替换为 packed 参数
Marlin fast path 仅接受 packed 权重；未预打包则 fail fast，避免运行时 packing 峰值/不确定性
更新 Marlin wrapper 的入参与 shape 推导（从 packed layout 计算 K/N）

区分pcie和hglink custom allreduce的使用
vllm：export VLLM_CUSTOM_CACHE=1
dtk：export HIP_KERNEL_EVENT_SYSTENFENCE=1

This reverts commit bed8bb81.

[Bugfix] Fixing trying to import non-existent symbols from libnccl.so

See merge request dcutoolkit/deeplearing/vllm!355

add torchvision deps

V1 采样器：新增 reduced top-k/top-p 采样路径

See merge request dcutoolkit/deeplearing/vllm!350

perf(fused-moe): 接入 W16A16 Marlin MoE 并缓存 pack 权重

See merge request dcutoolkit/deeplearing/vllm!347