Skip to main content

​
GPU Selection Commands

Run your code on multiple GPUs to determine optimal hardware based on memory requirements, runtime performance, and cost.

​
Commands

CommandDescription
lyceum gpu-selection runAnalyze code and find the optimal GPU
lyceum gpu-selection statusCheck status of a GPU selection job

​
lyceum gpu-selection run

Submits your code to run on all GPU profiles available to your account, then returns which GPU performed best. 
lyceum gpu-selection run <code_or_file>

​
Arguments

ArgumentDescription
code_or_file(required) Python code to execute or path to a Python file

​
Options

OptionDescription
--file-name, -fName for the execution
--timeout, -tTimeout per sub-job in seconds (1-600). Default: 60
--requirements, -rRequirements file path or pip requirements string
--importPre-import modules (can be used multiple times)
--use-config/--no-configUse workspace config from .lyceum/config.json if available. Default: enabled
--optimize, -oOptimization strategy: cost (cheapest), speed (fastest), or util (highest utilization). Default uses API recommendation
--debug, -dShow detailed debug information

​
Script Arguments

Pass arguments to your script using -- separator: 
lyceum gpu-selection run train.py -- --epochs 10 --lr 0.001

​
Examples

# Analyze a training script
lyceum gpu-selection run train.py

# Analyze with custom timeout
lyceum gpu-selection run train.py --timeout 120

# Optimize for cost
lyceum gpu-selection run train.py --optimize cost

# Optimize for speed
lyceum gpu-selection run train.py --optimize speed

# Pass script arguments
lyceum gpu-selection run train.py -- --batch-size 32 --epochs 5

​
Output

The command provides:
  • Memory Breakdown: Shows parameter count, model weights, gradients, optimizer states, and activations
  • Compatible GPUs: Lists all GPUs that can run your workload with VRAM requirements and utilization
  • Runtime Prediction: Estimated execution time and cost for each GPU
  • Recommendation: The best GPU based on your optimization strategy
After analysis, you’ll see a suggested run command: 
▶ Run on Lyceum:

  lyceum python run train.py -m gpu.a100

​
lyceum gpu-selection status

Check the status of a previously submitted GPU selection job. 
lyceum gpu-selection status <execution_id>

​
Arguments

ArgumentDescription
execution_id(required) Execution ID to check

​
Options

OptionDescription
--optimize, -oOptimization strategy for displaying results: cost, speed, or util

​
Examples

# Check job status
lyceum gpu-selection status exec_abc123

# Check status optimizing for cost
lyceum gpu-selection status exec_abc123 --optimize cost

​
Supported GPUs

GPU selection analyzes performance across these GPU types:
GPUVRAMUse Case
T416 GBDevelopment, small models
L40S48 GBMedium models, inference
A100 (40GB)40 GBTraining, large models
A100 (80GB)80 GBLarge models, multi-GPU
H10080 GBFastest training, LLMs
H200141 GBLargest models

​
Requirements

Your code must use PyTorch or the HuggingFace ecosystem for GPU selection to analyze memory and performance requirements. The analysis works best with:
  • Models defined as nn.Module subclasses
  • GPU operations using .to('cuda') or device = torch.device('cuda')
  • Training loops with loss.backward() and optimizer.step()