77 lines
2.4 KiB
Text
77 lines
2.4 KiB
Text
backends:
|
|
anthropic:
|
|
api_key_env: ANTHROPIC_API_KEY
|
|
enabled: false
|
|
model: claude-sonnet-4-6
|
|
type: anthropic
|
|
supports_images: true
|
|
claude_code:
|
|
api_key: any
|
|
base_url: http://localhost:3009/v1
|
|
enabled: false
|
|
model: claude-code-terminal
|
|
type: openai_compat
|
|
supports_images: true
|
|
github_copilot:
|
|
api_key: any
|
|
base_url: http://localhost:3010/v1
|
|
enabled: false
|
|
model: gpt-4o
|
|
type: openai_compat
|
|
supports_images: false
|
|
ollama:
|
|
api_key: ollama
|
|
base_url: http://ollama:11434/v1 # Docker service name; use localhost:11434 outside Docker
|
|
enabled: true
|
|
model: llama3.2:3b
|
|
type: openai_compat
|
|
supports_images: false
|
|
ollama_research:
|
|
api_key: ollama
|
|
base_url: http://ollama:11434/v1 # Docker service name; use localhost:11434 outside Docker
|
|
enabled: true
|
|
model: llama3.2:3b
|
|
type: openai_compat
|
|
supports_images: false
|
|
vllm:
|
|
api_key: ''
|
|
base_url: http://vllm:8000/v1 # Docker service name; use localhost:8000 outside Docker
|
|
enabled: true
|
|
model: __auto__
|
|
type: openai_compat
|
|
supports_images: false
|
|
vision_service:
|
|
base_url: http://localhost:8002
|
|
enabled: false
|
|
type: vision_service
|
|
supports_images: true
|
|
fallback_order:
|
|
- ollama
|
|
- claude_code
|
|
- vllm
|
|
- github_copilot
|
|
- anthropic
|
|
research_fallback_order:
|
|
- claude_code
|
|
- vllm
|
|
- ollama_research
|
|
- github_copilot
|
|
- anthropic
|
|
vision_fallback_order:
|
|
- vision_service
|
|
- claude_code
|
|
- anthropic
|
|
# Note: 'ollama' (alex-cover-writer) intentionally excluded — research
|
|
# must never use the fine-tuned writer model, and this also avoids evicting
|
|
# the writer from GPU memory while a cover letter task is in flight.
|
|
|
|
# ── Scheduler — LLM batch queue optimizer ─────────────────────────────────────
|
|
# The scheduler batches LLM tasks by model type to avoid GPU model switching.
|
|
# VRAM budgets are conservative peak estimates (GB) for each task type.
|
|
# Increase if your models are larger; decrease if tasks share GPU memory well.
|
|
scheduler:
|
|
vram_budgets:
|
|
cover_letter: 2.5 # alex-cover-writer:latest (~2GB GGUF + headroom)
|
|
company_research: 5.0 # llama3.1:8b or vllm model
|
|
wizard_generate: 2.5 # same model family as cover_letter
|
|
max_queue_depth: 500 # max pending tasks per type before drops (with logged warning)
|