tfmx 1.1.4

Serve Transformers and Manage GPUs

tfmx

Install

pip install tfmx --upgrade

Commands

Set GPU control state:

gpu_fan -cs a:1

Set GPU power limit:

# M-X GPU-0/1
gpu_pow -pm a:1 && gpu_pow -pl "0:160;1:200"

# M-A GPU-0/1
gpu_pow -pm a:1 && gpu_pow -pl "0,1:160"

Set GPU fan speed:

gpu_fan -cs a:1 && gpu_fan -fs a:100

Set GPU monitor with curve:

# gpu_mon -c "a:30-50/50-65/60-80/75-100;3,7:25-100" -s

Run tei compose and machine:

tei machine run --auto-start --perf-track --on-conflict replace

Run tei benchmark:

tei benchmark run -E "http://localhost:28800" -n 100000

Run qwn compose, machine, and benchmark:

export QWN_MACHINE_URL="http://$QWN_HOST:27800"

qwn compose up --gpu-configs "0"
qwn compose up --gpu-layout uniform-awq
qwn machine run --auto-start -b --on-conflict replace
qwn client chat "你好，请做个自我介绍"
qwn benchmark run -E "$QWN_MACHINE_URL" -n 100

Staged Run Scripts

For repeatable repo-local workflows, use the staged script directories:

bash runs/teis/01_deploy_default.sh
bash runs/teis/02_start_machine.sh
bash runs/teis/03_health_check.sh
bash runs/qwns/01_deploy_uniform.sh
bash runs/qwns/02_start_machine.sh
bash runs/qwns/03_health_check.sh
bash runs/recovery/restart_tei_qwn.sh

• runs/teis/README.md: staged TEI deploy, health, benchmark, cleanup
• runs/qwns/README.md: staged QWN deploy, health, benchmark, cleanup
• runs/recovery/README.md: joint TEI + QWN recovery and live validation