API Updates: routing table, models endpoint, inference routing, rebase on safety_refactor

[yanxi0830]

Changes

Major Changes

✅ rebase on top of safety_refactor branch
✅ routers migration
- move memory router to use llama_stack/distribution/routers
- add RoutingTable class with to manage provider implementations
- add inference routing based on model param to use different models
✅ models/list & models/get endpoint to query information on what model is currently being served & corresponding api provider
- 1 adapter could support multiple models (e.g. fireworks/together/ollama), for this case, we list all supported models by the adapter
✅ safety/list_shields to query available shields in distribution

Minor Changes

✅ backward compatibility: single adapter without routing implementation still supported
✅ example configs to configure advanced routing table
✅ configure script to configure single adapter backed endpoints

Not in this PR

🚧 register_model endpoint to add new models
🚧 Client CLI for querying distribution models endpoint --> will be in llama-stack-apps
🚧 Update client SDK package to deal with new OpenAPI spec on run_shield

Tests

Routing Table

Test Inference Routing

python -m llama_stack.apis.inference.client 
python sdk_examples/inference/client.py localhost 5000 false

• Meta-Llama3.1-8B-Instruct
  
• Meta-Llama3.1-8B
  

Test Memory Routing

python -m llama_stack.apis.memory.client

• vector memory type
  

• switch to keyvalue memory type remote::pgvector
  

Test Agents

python -m llama_stack.apis.agents.client
python sdk_examples/agents/client.py

Test Safety

python -m llama_stack.apis.safety.client localhost 5000

Models Endpoint with Routing Table

• using ollama supported models list

inference:
    provider_id: remote::ollama
    config:
      url: https://xxx

• using default local

inference:
    provider_id: meta-reference
    config:
      model: Meta-Llama3.1-8B-Instruct
      quantization: null
      torch_seed: null
      max_seq_len: 4096
      max_batch_size: 1

Models Endpoint without Routing Table

• using routing table

provider_routing_table:
  inference:
    - routing_key: Meta-Llama3.1-8B-Instruct
      provider_id: meta-reference
      config:
        model: Meta-Llama3.1-8B-Instruct
        quantization: null
        torch_seed: null
        max_seq_len: 4096
        max_batch_size: 1
    - routing_key: Meta-Llama3.1-8B
      provider_id: remote::ollama
      config:
        url: https:://ollama.com

run.yaml

built_at: '2024-09-18T13:41:17.656743'
image_name: local
docker_image: null
conda_env: local
apis_to_serve:
- inference
- memory
- telemetry
- agents
- safety
provider_map:
  telemetry:
    provider_id: meta-reference
    config: {}
  safety:
    provider_id: meta-reference
    config:
      llama_guard_shield:
        model: Llama-Guard-3-8B
        excluded_categories: []
        disable_input_check: false
        disable_output_check: false
      prompt_guard_shield:
        model: Prompt-Guard-86M
  agents:
    provider_id: meta-reference
    config: {}
provider_routing_table:
  inference:
    - routing_key: Meta-Llama3.1-8B-Instruct
      provider_id: meta-reference
      config:
        model: Meta-Llama3.1-8B-Instruct
        quantization: null
        torch_seed: null
        max_seq_len: 4096
        max_batch_size: 1
    # - routing_key: Meta-Llama3.1-8B
    #   provider_id: meta-reference
    #   config:
    #     model: Meta-Llama3.1-8B
    #     quantization: null
    #     torch_seed: null
    #     max_seq_len: 4096
    #     max_batch_size: 1
  memory:
    # - routing_key: keyvalue
    #   provider_id: remote::pgvector
    #   config:
    #     host: localhost
    #     port: 5432
    #     db: vectordb
    #     user: vectoruser
    #     password: xxxx
    - routing_key: vector
      provider_id: meta-reference
      config: {}