Command Line Reference
This document provides a comprehensive reference for all command line interfaces in TensorFusion.
| Parameter | Description | Default |
|---|
-enable-http2 | Enables HTTP/2 for the metrics and webhook servers | - |
-health-probe-bind-address | The address the probe endpoint binds to | :8081 |
-kubeconfig | Path to a kubeconfig file (only required if out-of-cluster) | - |
-leader-elect | Enable leader election for controller manager to ensure only one active instance | - |
-metrics-bind-address | The address the metrics endpoint binds to | 0 (disabled) |
-metrics-secure | Serve metrics endpoint securely via HTTPS (use --metrics-secure=false for HTTP) | - |
-zap-devel | Use development mode for logging | true |
-zap-encoder | Zap log encoding format (json or console) | - |
-zap-log-level | Verbosity level for logging (debug, info, error, or any integer value > 0) | - |
-zap-stacktrace-level | Level at which stacktraces are captured (info, error, panic) | - |
-zap-time-encoding | Time encoding format (epoch, millis, nano, iso8601, rfc3339, rfc3339nano) | epoch |
| Variable | Description | Example |
|---|
INITIAL_GPU_NODE_LABEL_SELECTOR | Initial label selector for GPU nodes | nvidia.com/gpu.present=true |
ENABLE_WEBHOOKS | Enable webhook functionality | true |
OPERATOR_NAMESPACE | Namespace for the operator | tensor-fusion-sys |
KUBECONFIG | Path to kubeconfig file | <kubeconfig> |
| Parameter | Description | Default |
|---|
--sock_path | Worker unix socket path | /tensor-fusion/worker/sock |
--gpu_metrics_file | GPU metrics file location | /logs/metrics.log |
--scheduler | Scheduling policy for multiple processes on single GPU node (when GPU load is high) | Options: FIFO for simple first-in-first-out, MLFQ for multi-level feedback queue |
| Parameter | Description | Example |
|---|
--hostname | Custom hostname for binding current node with GPUNode custom resource | <hostname> |
--gpu-info-config | Path to the GPU info configuration file | See example below |
- model: RTX5090
fullModelName: "NVIDIA GeForce RTX 5090"
vendor: NVIDIA
costPerHour: 0.65
fp16TFlops: 419
| Variable | Description | Example |
|---|
HOSTNAME | Node hostname | <hostname> |
KUBECONFIG | Path to kubeconfig file | <kubeconfig> |
NODE_DISCOVERY_REPORT_GPU_NODE | GPU node custom resource name | <gpu-node-custom-resource-name> |
| Parameter | Description | Default/Notes |
|---|
-n | Network protocol | Currently only native (native TCP communication) |
-p | Worker port | Random value assigned by TensorFusion Operator-Scheduler |
-s | Unix socket path folder | Should be /tensor-fusion/worker/sock/ in Kubernetes |
| Variable | Description | Value |
|---|
TF_ENABLE_LOG | Enable logging | 1 |
The GPU Client Stub consists of two libraries that use LD_PRELOAD to run before every process started inside the container or server:
libadd_path.so: Adds additional library paths for AI application environments (e.g., hooked NVML)libcuda.so: Hooks into CUDA runtime
Example configuration in worker template:
env:
- name: LD_PRELOAD
value: /tensor-fusion/libadd_path.so:/tensor-fusion/libcuda.so
| Variable | Description | Value/Notes |
|---|
TF_PATH | Appended to PATH environment variable | /tensor-fusion |
TF_LD_PRELOAD | Appended to LD_PRELOAD | Varies |
TF_LD_LIBRARY_PATH | Appended to LD_LIBRARY_PATH | /tensor-fusion |
TF_ENABLE_LOG | Disable/Enable logging, default to disabled | 0 |
TensorFusion Docs