Stop here if any apply:

Past this point, you are fighting physics, not software. VRAM is a fixed resource. No config change can make 8GB physically hold 20GB.

Use when: Your GPU has sufficient VRAM for the model, or you're willing to use a smaller model.

Why this works:

• Quantization reduces VRAM usage by 60-75%
• Context reduction limits KV cache growth
• Model selection matches VRAM to requirements

Time investment: 10-15 minutes

Solution 1: Use a Quantized Model (Recommended)

Quantized models use fewer bits per parameter, dramatically reducing VRAM requirements.

# Check available quantized versions
ollama list | grep deepseek

# Run Q4 quantized version (uses ~60% less VRAM)
ollama run deepseek-r1:32b-q4_K_M

# Or use the 8B model instead (much lower VRAM)
ollama run deepseek-r1:8b-q4_K_M

# Configure OpenClaw to use quantized model
export OPENCLAW_MODEL="deepseek-r1:32b-q4_K_M"
openclaw serve

Solution 2: Reduce Context Window

Limiting the context window reduces KV cache size and prevents Agent Loop OOM.

# Default is usually 16384, cut to 4096 or 2048
ollama run deepseek-r1:32b-q4_K_M \
  --num_ctx 4096 \
  --num-gpu 99 \
  --repeat-penalty 0.6

Solution 3: Partial GPU Offload

Let GPU handle some layers, CPU handles the rest. Slower, but uses less VRAM.

# Only load 35 layers on GPU, rest on CPU
ollama run deepseek-r1:32b-q4_K_M \
  --num-gpu-layers 35 \
  --num_ctx 2048

# Warning: CPU inference is 5-10x slower

Solution 4: Enable Conversation Pruning

OpenClaw can automatically drop old messages to prevent KV cache overflow.

# openclaw_config.py
from openclaw import Client

client = Client(model="deepseek-r1:8b-q4_K_M")

# Auto-prune conversation history
client.max_history_messages = 10
client.context_window = 4096

# OpenClaw drops old messages to prevent OOM
response = client.generate("Your prompt here")

Verification:

# Monitor VRAM while running
watch -n 1 nvidia-smi

# In another terminal, run OpenClaw
# VRAM should stabilize below 100%

Use when: Primary Exit Path fails — your GPU physically cannot hold the model.

This is the only remaining technically valid option when:

• Your GPU VRAM < 6GB (cannot run 8B even with quantization)
• You need DeepSeek R1 32B or 70B for your use case
• You require production reliability (consumer GPUs are not designed for 24/7 workloads)
• You've quantized, reduced context, and still hit OOM

Solution: Cloud GPU with Dedicated VRAM

Cloud GPUs provide the VRAM needed for larger models. This is not a "workaround" — it's the correct solution when hardware constraints are absolute.

Why this is valid:

• VRAM is a physical resource. You cannot add VRAM to a consumer GPU.
• Cloud GPUs (H100, A100) have 40GB-80GB VRAM — sufficient for any DeepSeek R1 variant.
• Hourly billing lets you test and scale.
• Production isolation prevents AI workloads from affecting your development machine.

Steps:

1. Choose a GPU provider with NVIDIA H100 or A100 availability
2. Select GPU VRAM based on your model:
   • DeepSeek R1 8B: 16GB+ VRAM
   • DeepSeek R1 32B: 24GB+ VRAM
   • DeepSeek R1 70B: 48GB+ VRAM
3. Install Ollama on the VPS
4. Run DeepSeek R1 at full precision (no quantization needed)

Time investment: 20-30 minutes (including VPS setup)

Note: This is not "giving up" on local debugging. It's recognizing that VRAM is a fixed resource and selecting the appropriate hardware for your workload.