System Info

System Info

Gather a complete picture of available compute before designing or launching experiments. After collecting the data, produce a structured profile and write it to memory/system-profile.md in the project sandbox.

---

When to Use

• At the start of any project involving training, fine-tuning, or large-scale data processing
• When a training run is unexpectedly slow or crashes (re-run to check resource pressure)
• Before deciding batch size, number of workers, mixed precision settings, or dataset size

---

Step 1 — Detect OS

uname -s

• Linux → use Linux commands below
• Darwin → use macOS commands below
• Windows → use PowerShell commands below (limited support)

---

Step 2 — CPU

Linux:

lscpu | grep -E "Model name|Socket|Core|Thread|MHz"
nproc

macOS:

sysctl -n machdep.cpu.brand_string
sysctl -n hw.physicalcpu hw.logicalcpu

Record: model name, physical cores, logical cores (threads).

---

Step 3 — RAM

Linux:

free -h

macOS:

sysctl -n hw.memsize | awk '{print $1/1024/1024/1024 " GB"}'
vm_stat | grep -E "Pages free|Pages active|Pages inactive|Pages wired"

Record: total RAM, available RAM. On macOS, note that vm_stat pages are 16KB each.

---

Step 4 — Disk

Linux / macOS:

df -h .

Run from within the project sandbox_root to see available space on the relevant partition.

Record: total disk, used, available on the partition where the project lives.

---

Step 5 — GPU Detection

Check for NVIDIA GPU (Linux / Windows WSL)

nvidia-smi --query-gpu=name,memory.total,memory.free,utilization.gpu,driver_version --format=csv,noheader

If nvidia-smi is not found, no NVIDIA GPU is available.

Also check CUDA availability via Python:

python3 -c "import torch; print('CUDA:', torch.cuda.is_available()); print('Device count:', torch.cuda.device_count()); [print(f'  GPU {i}:', torch.cuda.get_device_name(i)) for i in range(torch.cuda.device_count())]"

Check for Apple Silicon MPS (macOS)

python3 -c "import torch; print('MPS available:', torch.backends.mps.is_available()); print('MPS built:', torch.backends.mps.is_built())"

Also check chip type:

sysctl -n machdep.cpu.brand_string   # shows "Apple M1/M2/M3/M4"
system_profiler SPHardwareDataType | grep "Chip"

Fallback — CPU only

If neither CUDA nor MPS is available, record compute: cpu-only.

---

Step 6 — Python and Deep Learning Environment

python3 --version
python3 -c "import torch; print('PyTorch:', torch.__version__)"
python3 -c "import torch; print('cuDNN:', torch.backends.cudnn.version())" 2>/dev/null || echo "cuDNN: not available"

If torch is not installed, note it — the project setup step will need to install it.

---

Step 7 — Write System Profile

Write findings to memory/system-profile.md inside the project sandbox. Use this template:

# System Profile

- **Collected**: 
- **OS**: 

## CPU
- **Model**: 
- **Physical cores**: 
- **Logical cores**: 

## RAM
- **Total**: 
- **Available**: 

## Disk (project partition)
- **Total**: 
- **Available**: 

## Compute
- **Type**: 
- **Device**: 
- **VRAM / Unified memory**: 
- **CUDA version**: 
- **MPS available**: 

## Python Environment
- **Python**: 
- **PyTorch**: 
- **cuDNN**: 

## Experiment Implications

---

Step 8 — Derive Experiment Implications

After collecting the profile, reason about what it means for the project goal. Add an ## Experiment Implications section to memory/system-profile.md. Consider:

Batch size:

• NVIDIA GPU with ≥40GB VRAM → large batches (512+) viable
• NVIDIA GPU with <16GB VRAM or MPS → moderate batches (64–256), watch for OOM
• CPU only → small batches (8–32), training will be slow; flag this clearly

Mixed precision:

• NVIDIA Ampere (A100, RTX 30xx/40xx) or newer → torch.float16 or bfloat16 safe
• Apple MPS → torch.float16 supported; bfloat16 not supported on all models
• CPU → keep float32

Number of dataloader workers:

• Set to min(physical_cores - 1, 8) as a safe default

Disk:

• If available disk < 10 GB, warn — checkpoints and datasets may fill it
• Note available space explicitly so the user can plan

Estimated training time:

• If CPU only and project involves a non-trivial model, flag that training may take hours and recommend checkpointing aggressively (every N steps, not just per epoch)

Example output:

## Experiment Implications
- **Compute**: MPS (Apple M3 Pro). Use `device = "mps"`. Mixed precision not recommended.
- **Batch size**: Start with 64. Increase if memory allows; watch `mps` backend for silent OOM.
- **Workers**: 4 (physical cores: 6, leaving 2 for system)
- **Disk**: 420 GB available — sufficient for checkpoints and shakespeare_char dataset.
- **Estimated time**: shakespeare_char on MPS ~15–30 min. CPU fallback ~2–4h.
- **Recommendation**: Checkpoint every 500 iterations given MPS instability on long runs.

---

Output

| File | Description | |---|---| | memory/system-profile.md | Full system profile + experiment implications |

This file is read by other agents (executor, orchestrator) when configuring training runs. It does not need to be included in every checkpoint — write it once at project start and update it only if hardware changes.

---

Credit: originally sourced from clawhub.ai

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: frankaging
• Source: frankaging/lite-research-agents
• License: MIT

Install and usage instructions live in the source repository linked above.