allsds

如何在2张C500部署 MiniMax-H3 量化8模型

根据github.com/MetaX-MACA/mxFIO 在c500 上测试gds 发现实际路径并没有实现 gpu 到storage
测试命令如下：

./fio  --name=gds-test  --directory=/mnt/gds/ext4  --ioengine=libcufile  --cuda_io=cufile  --gpu_dev_ids=0   --direct=1  --rw=write   --bs=1M  --size=4G  --numjobs=1  --group_reporting   --time_based   --runtime=60 --ramp_time=4 --end_fsync=1

mcFile 日志发现回退到cpu copy ，是什么原因呢？

switch-PURICO:21057:21057 [MCFILE DEBUG] (mcFileWrite:217) API_CALL devPtr_base = 0x74da76600000, size = 1048576, file_offset = 1115684864, devPtr_offset = 0
switch-PURICO:21057:21057 [MCFILE DEBUG] (write:150) ioctl<MXCD_IOC_VRAM_IO> failed, ret = -1, errno = 22, args.size = 1048576
switch-PURICO:21057:21057 [MCFILE DEBUG] (write:161) direct copy failed, fallback to cpu copy
switch-PURICO:21057:21057 [MCFILE DEBUG] (write:182) size = 1048576, dur = 465usec(s), speed = 2.10GB/sec
switch-PURICO:21057:21057 [MCFILE DEBUG] (write:184) Host
switch-PURICO:21057:21057 [MCFILE DEBUG] (mcFileWrite:217) API_CALL devPtr_base = 0x74da76600000, size = 1048576, file_offset = 1116733440, devPtr_offset = 0
switch-PURICO:21057:21057 [MCFILE DEBUG] (write:150) ioctl<MXCD_IOC_VRAM_IO> failed, ret = -1, errno = 22, args.size = 1048576
switch-PURICO:21057:21057 [MCFILE DEBUG] (write:161) direct copy failed, fallback to cpu copy
switch-PURICO:21057:21057 [MCFILE DEBUG] (write:182) size = 1048576, dur = 457usec(s), speed = 2.14GB/sec
switch-PURICO:21057:21057 [MCFILE DEBUG] (write:184) Host

mx-smi 如下