GX10 Offload

Offload work to the local NVIDIA DGX Spark cluster node running Ollama with Devstral models on GB10 Blackwell GPU (128GB unified memory).

When to Use

• Long code generation tasks that benefit from a dedicated local model
• Batch processing of multiple prompts
• Draft generation for review (speculative decoding pattern)
• Tasks where latency to cloud APIs is a bottleneck
• Privacy-sensitive inference that must stay on-premises

Connection

10.0.0.234

gx10-94e2.local

100.67.53.87

a

aaaaaa

http://localhost:11434

ssh -L 11434:localhost:11434 a@10.0.0.234

Note: The WiFi-connected unit is gx10-94e2 (discovered via mDNS). The Tailscale-reachable unit is gx10-acee (different physical Spark).

Available Models

devstral

devstral-2:123b

devstral2-4k

Quick Usage

Single prompt via SSH

sshpass -p 'aaaaaa' ssh -o PreferredAuthentications=password -o PubkeyAuthentication=no \
  a@100.67.53.87 \
  "curl -s http://localhost:11434/api/generate -d '{\"model\":\"devstral\",\"prompt\":\"YOUR_PROMPT\",\"stream\":false}'"

Via SSH tunnel (persistent)

# Open tunnel in background
sshpass -p 'aaaaaa' ssh -o PreferredAuthentications=password -o PubkeyAuthentication=no \
  -fNL 11434:localhost:11434 a@100.67.53.87

# Then use locally as if Ollama were running here
curl http://localhost:11434/api/generate \
  -d '{"model":"devstral","prompt":"Hello","stream":false}'

OpenAI-compatible API

curl http://localhost:11434/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "devstral",
    "messages": [{"role": "user", "content": "Write a Python function to sort a list"}]
  }'

Using the offload script

# Simple prompt
~/.claude/skills/gx10-offload/scripts/offload.sh "Write a Rust function for binary search"

# With specific model
~/.claude/skills/gx10-offload/scripts/offload.sh "Explain monads" devstral-2:123b

# Batch mode (one prompt per line)
~/.claude/skills/gx10-offload/scripts/offload.sh --batch prompts.txt

Offload Patterns

1. Draft-and-Review

Offload draft generation to GX10, then review/refine with Claude:

# GX10 generates draft
DRAFT=$(~/.claude/skills/gx10-offload/scripts/offload.sh "Implement a Redis cache wrapper in Python with TTL support")
# Claude reviews and improves the draft

2. Batch Code Generation

Generate multiple implementations in parallel on GX10:

for task in "sort" "search" "hash" "tree"; do
  ~/.claude/skills/gx10-offload/scripts/offload.sh "Implement $task in Rust" &
done
wait

3. Test Generation

Offload test writing to the local model:

~/.claude/skills/gx10-offload/scripts/offload.sh "Write pytest tests for: $(cat src/main.py)"

Device Status Check

~/.claude/skills/gx10-offload/scripts/offload.sh --status

Ensure Ollama is Running

sshpass -p 'aaaaaa' ssh -o PreferredAuthentications=password -o PubkeyAuthentication=no \
  a@100.67.53.87 'pgrep ollama || nohup ollama serve > /tmp/ollama.log 2>&1 &'

Hardware

• GPU: NVIDIA GB10 Blackwell (DGX Spark)
• Memory: 128GB unified (Grace-Blackwell architecture)
• CPU: 20-core Grace ARM64
• OS: Ubuntu 24.04 aarch64, kernel 6.14-nvidia
• PyTorch: 2.10.0 with CUDA
• Disk: 510GB free

GF(3) Assignment

Conservation triad:

gx10-offload (+1) + tailscale (0) + skill-creator (-1) = 0