mirror of
https://github.com/open-jarvis/OpenJarvis.git
synced 2026-08-14 00:47:52 +00:00
115 lines
5.1 KiB
TOML
115 lines
5.1 KiB
TOML
# OpenJarvis configuration — GLM-4.7-Flash eval on 8x A100-80GB
|
|
|
|
# ═══════════════════════════════════════════════════════════════
|
|
# PILLAR 1: Intelligence — The Model
|
|
# ═══════════════════════════════════════════════════════════════
|
|
[intelligence]
|
|
default_model = "zai-org/GLM-4.7-Flash" # HuggingFace model ID
|
|
fallback_model = "glm-4.7-flash" # Catalog alias fallback
|
|
preferred_engine = "vllm" # MoE model, vLLM is best for A100s
|
|
provider = "local" # Running locally on this machine
|
|
quantization = "none" # Full precision, we have 640GB VRAM
|
|
# Generation defaults for eval
|
|
temperature = 0.0 # Deterministic for reproducibility
|
|
max_tokens = 2048 # Standard eval output length
|
|
top_p = 0.9
|
|
top_k = 40
|
|
repetition_penalty = 1.0
|
|
|
|
# ═══════════════════════════════════════════════════════════════
|
|
# PILLAR 2: Agent — The Agentic Harness
|
|
# ═══════════════════════════════════════════════════════════════
|
|
[agent]
|
|
default_agent = "native_openhands" # CodeAct-style agent
|
|
max_turns = 10 # Up to 10 tool-calling turns
|
|
tools = "code_interpreter,web_search,file_read,calculator,think"
|
|
objective = "Answer questions accurately using available tools"
|
|
context_from_memory = false # No memory injection during eval
|
|
|
|
# ═══════════════════════════════════════════════════════════════
|
|
# PILLAR 3: Tools — MCP Interface
|
|
# ═══════════════════════════════════════════════════════════════
|
|
[tools.storage]
|
|
default_backend = "sqlite"
|
|
db_path = "~/.openjarvis/memory.db"
|
|
|
|
[tools.mcp]
|
|
enabled = true
|
|
|
|
# ═══════════════════════════════════════════════════════════════
|
|
# PILLAR 4: Engine — The Inference Runtime
|
|
# ═══════════════════════════════════════════════════════════════
|
|
[engine]
|
|
default = "vllm"
|
|
|
|
[engine.vllm]
|
|
host = "http://localhost:8001" # vLLM serving port
|
|
|
|
[engine.ollama]
|
|
host = "http://localhost:11434"
|
|
|
|
[engine.sglang]
|
|
host = "http://localhost:30000"
|
|
|
|
[engine.llamacpp]
|
|
host = "http://localhost:8080"
|
|
|
|
[engine.exo]
|
|
host = "http://localhost:52415"
|
|
|
|
[engine.nexa]
|
|
host = "http://localhost:18181"
|
|
# device = "npu" # optional: cpu, gpu, npu
|
|
|
|
[engine.uzu]
|
|
host = "http://localhost:8080"
|
|
|
|
[engine.apple_fm]
|
|
host = "http://localhost:8079"
|
|
|
|
# ═══════════════════════════════════════════════════════════════
|
|
# PILLAR 5: Learning — Improvement Methodologies
|
|
# ═══════════════════════════════════════════════════════════════
|
|
[learning]
|
|
enabled = false # No learning during eval
|
|
|
|
[learning.routing]
|
|
policy = "heuristic"
|
|
|
|
[learning.intelligence]
|
|
policy = "none"
|
|
|
|
[learning.agent]
|
|
policy = "none"
|
|
|
|
[learning.metrics]
|
|
accuracy_weight = 0.6
|
|
latency_weight = 0.2
|
|
cost_weight = 0.1
|
|
efficiency_weight = 0.1
|
|
|
|
# ═══════════════════════════════════════════════════════════════
|
|
# Supporting config
|
|
# ═══════════════════════════════════════════════════════════════
|
|
[telemetry]
|
|
enabled = true
|
|
db_path = "~/.openjarvis/telemetry.db"
|
|
gpu_metrics = true
|
|
gpu_poll_interval_ms = 50
|
|
energy_vendor = "" # Auto-detect; or force "nvidia"/"amd"/"apple"/"cpu_rapl"
|
|
warmup_samples = 0 # Warmup iterations before steady-state measurement
|
|
steady_state_window = 5 # Sliding window size for CV stability check
|
|
steady_state_threshold = 0.05 # Coefficient of variation threshold for steady state
|
|
|
|
[traces]
|
|
enabled = true # Record traces for analysis
|
|
db_path = "~/.openjarvis/traces.db"
|
|
|
|
[server]
|
|
host = "0.0.0.0"
|
|
port = 8000
|
|
agent = "native_openhands"
|
|
|
|
[security]
|
|
enabled = false # Disable for eval (no PII scanning overhead)
|