Project-AI

DeepSeek V3.2 Integration Guide

Overview

This guide covers the integration of DeepSeek V3.2 Mixture-of-Experts (MoE) language model into Project-AI. The integration provides a modular, production-ready interface for advanced language generation tasks.

Quick Start

Installation

# Install dependencies
pip install -r requirements.txt

# Optional: Set Hugging Face API key in .env (for API access)
echo "HUGGINGFACE_API_KEY=your_key_here" >> .env

Basic Usage

Python API

from app.core.deepseek_v32_inference import DeepSeekV32

# Initialize (auto-detects GPU/CPU)
deepseek = DeepSeekV32()

# Text completion
result = deepseek.generate_completion(
    prompt="Explain quantum computing",
    max_new_tokens=256,
    temperature=0.7
)
print(result["text"])

# Chat mode
messages = [
    {"role": "user", "content": "What is machine learning?"}
]
result = deepseek.generate_chat(messages)
print(result["text"])

Command-Line Interface

# Simple completion
python scripts/deepseek_v32_cli.py "Explain artificial intelligence"

# Chat mode
python scripts/deepseek_v32_cli.py --mode chat "Hello, how are you?"

# Interactive chat session
python scripts/deepseek_v32_cli.py --mode chat --interactive

# Custom parameters
python scripts/deepseek_v32_cli.py \
  --temperature 0.8 \
  --max-tokens 512 \
  --device cuda \
  "Your prompt here"

Features

1. Dual Inference Modes

Completion Mode: Generate text continuations

result = deepseek.generate_completion("Once upon a time")

Chat Mode: Multi-turn conversations

messages = [
    {"role": "user", "content": "Hello"},
    {"role": "assistant", "content": "Hi! How can I help?"},
    {"role": "user", "content": "Tell me about AI"}
]
result = deepseek.generate_chat(messages)

2. Content Safety

Built-in content filtering blocks inappropriate prompts:

# Safe prompt - passes
result = deepseek.generate_completion("Explain machine learning")

# Unsafe prompt - blocked
result = deepseek.generate_completion("Generate illegal content")
# Returns: {"success": False, "error": "Content filter: Blocked keyword detected: illegal"}

Disable filtering (use with caution):

deepseek.content_filter_enabled = False

3. GPU Acceleration

Automatic device detection with fallback:

# Auto-detect (prefers CUDA > MPS > CPU)
deepseek = DeepSeekV32()

# Force specific device
deepseek = DeepSeekV32(device="cuda")  # NVIDIA GPU
deepseek = DeepSeekV32(device="mps")   # Apple Silicon
deepseek = DeepSeekV32(device="cpu")   # CPU fallback

4. Parameter Tuning

Fine-tune generation behavior:

# At initialization
deepseek = DeepSeekV32(
    temperature=0.8,    # Randomness (0.0-2.0)
    top_p=0.95,         # Nucleus sampling
    top_k=100,          # Top-k sampling
    max_length=512      # Max tokens
)

# Update at runtime
deepseek.update_parameters(
    temperature=0.9,
    max_length=1024
)

5. Memory Management

Efficient model loading/unloading:

# Load model (lazy loaded on first inference)
deepseek = DeepSeekV32()

# Unload to free memory
deepseek.unload_model()

Advanced Usage

Custom Model

Use a different model from Hugging Face:

deepseek = DeepSeekV32(model_name="custom/model-name")

Batch Processing

prompts = [
    "Explain AI",
    "What is quantum computing?",
    "Define machine learning"
]

for prompt in prompts:
    result = deepseek.generate_completion(prompt)
    print(f"Q: {prompt}")
    print(f"A: {result['text']}\n")

Error Handling

result = deepseek.generate_completion("Your prompt")

if result["success"]:
    print("Response:", result["text"])
else:
    print("Error:", result["error"])

Integration with Project-AI

With Four Laws AI Core

from app.core.ai_systems import FourLaws
from app.core.deepseek_v32_inference import DeepSeekV32

deepseek = DeepSeekV32()

# Validate action before execution
is_allowed, reason = FourLaws.validate_action(
    "Generate text",
    context={"is_user_order": True}
)

if is_allowed:
    result = deepseek.generate_completion("Your prompt")

With Memory System

from app.core.ai_systems import MemoryExpansionSystem
from app.core.deepseek_v32_inference import DeepSeekV32

memory = MemoryExpansionSystem()
deepseek = DeepSeekV32()

# Generate and store
result = deepseek.generate_completion("Explain AI")
if result["success"]:
    memory.add_knowledge("ai_responses", "ai_explanation", result["text"])

With Agent Council

# Use DeepSeek as reasoning engine for agents
from app.core.deepseek_v32_inference import DeepSeekV32

class ReasoningAgent:
    def __init__(self):
        self.deepseek = DeepSeekV32()
    
    def reason(self, task):
        prompt = f"Analyze this task and provide reasoning: {task}"
        result = self.deepseek.generate_completion(prompt)
        return result["text"]

Performance Tips

  1. GPU Acceleration: Use CUDA-enabled GPU for 10-100x speedup
  2. Model Caching: Keep model loaded for multiple inferences
  3. Batch Size: Process multiple prompts in sequence efficiently
  4. Parameter Tuning: Lower temperature for deterministic output
  5. Memory Management: Unload model when not in use for extended periods

Model Information

Troubleshooting

Model Download Issues

# Ensure sufficient disk space (40GB+)
df -h

# Check network connectivity to Hugging Face
curl -I https://huggingface.co/

Memory Issues

# Reduce max_length
deepseek = DeepSeekV32(max_length=256)

# Use CPU if GPU OOM
deepseek = DeepSeekV32(device="cpu")

# Unload model after use
deepseek.unload_model()

Import Errors

# Reinstall dependencies
pip install --force-reinstall transformers accelerate torch

Testing

Run comprehensive test suite:

# Unit tests
pytest tests/test_deepseek_v32.py -v

# Integration tests
pytest tests/test_deepseek_integration.py -v

# All tests
pytest tests/test_deepseek_*.py -v

Examples

See examples/deepseek_demo.py for complete working examples:

python examples/deepseek_demo.py

Support

License

Part of Project-AI, released under MIT License.

This site is open source. Improve this page.