跳转至

FAQ - Frequently Asked Questions

1. How to fix asyncio conflicts in Jupyter Notebook?

Issue: Running loop.run_until_complete in Notebooks causes RuntimeError: This event loop is already running. Cause: Jupyter already has a running event loop. Standard asyncio does not support nested loops by default. Solution: Install and use nest_asyncio to allow nested event loops:

import nest_asyncio
nest_asyncio.apply()

# Now you can run BEAM asynchronous methods (like agen) normally

2. How to configure API Keys?

Option A: Hardcoded in code (Best for quick experiments)

Pass the api_key directly into LLMRegistry.get():

llm = LLMRegistry.get(model_name="gpt-4o", api_key="sk-...")

Option B: Environment Variables

Set them in your system terminal or .env file. BEAM automatically scans for these if the parameter is missing:

  • OpenAI: export OPENAI_API_KEY='sk-...'
  • DeepSeek: export DEEPSEEK_API_KEY='sk-...'

3. How to handle CUDA Out of Memory (OOM) during training?

Solutions:

  1. Reduce Batch Size: Lower the batch_size in your optimizer configuration to reduce immediate VRAM usage.
  2. Decrease Initial Density: Lower initial_spatial_probability in OptimizationConfig to create a sparser initial graph.
  3. Gradient Accumulation: If a large batch is necessary, compute gradients over smaller steps before updating.
  4. CPU Execution: For purely structural/topological optimizations that aren't compute-heavy, set device='cpu'.

4. What if performance drops too much after Pruning?

Solutions:

  • Lower Pruning Rate: Set a smaller pruning_rate (e.g., 0.05 or 0.1) for a more gradual removal of edges.
  • Extend Warm-up Period: Increase the training rounds before triggering update_masks to ensure edge weights (logits) have converged correctly.
  • Refine Reward Function: If the reward signal is too sparse (mostly zeros), the algorithm might prune critical paths randomly due to lack of feedback.
  • Checkpoint Rollback: Use strategy.save() to save checkpoints regularly. If accuracy collapses, use strategy.load() to revert to the last stable topology.

Routing Logic:

BEAM’s LLMRegistry uses keyword-based routing based on the model_name:

  • Contains gpt -> Routes to OpenAILLM
  • Contains deepseek -> Routes to DeepSeekLLM (automatically sets DeepSeek endpoints)
  • Contains llama or qwen -> Routes to LocalLLM (requires base_url)