solemn ibex
So, background info. Been trying to have local hosted models run with Openclaw for ~ a month now, on an Oracle hosted VPS with 24 GB of ram.
However, been running into issues. I'll add more details below, but the gist is that time ollama run qwen 2.5:14b "reply with exactly: pong"
returns pretty quickly,
but
time openclaw agent --agent main --session-id pong-test-1 --local --message "Reply with exactly: PONG" --timeout 300 --json > /tmp/pong1.log 2>&1
tail -n 80 /tmp/pong1.log
does not.
gaunt gateBOT
solemn ibex
Openclaw call taking multiple times longer / not working with Ollama, when Ollama does not
Ollama:
ubuntu@openclaw-server:~$ time ollama run qwen2.5:14b "reply with exactly: pong"
pong
real 0m38.570s
user 0m0.166s
sys 0m0.124s
ubuntu@openclaw-server:~$ time ollama run qwen2.5:14b "reply with exactly: pong"
pong
real 0m1.325s
user 0m0.022s
sys 0m0.022s
openclaw:
time openclaw agent --agent main --session-id pong-test-1 --local --message "Reply with exactly: PONG" --timeout 900 --json > /tmp/pong1.log 2>&1
tail -n 80 /tmp/pong1.log
real 4m5.619s
user 1m39.324s
sys 0m45.174s
"missing": false,
"rawChars": 850,
"injectedChars": 0,
"truncated": true
},
{
"name": "IDENTITY.md",
"path": "/home/ubuntu/.openclaw/workspace/IDENTITY.md",
"missing": false,
"rawChars": 633,
"injectedChars": 0,
"truncated": true
},
{
"name": "USER.md",
"path": "/home/ubuntu/.openclaw/workspace/USER.md",
"missing": false,
"rawChars": 474,
"injectedChars": 0,
"truncated": true
},
{
"name": "BOOTSTRAP.md",
"path": "/home/ubuntu/.openclaw/workspace/BOOTSTRAP.md",
"missing": true,
"rawChars": 0,
"injectedChars": 0,
"truncated": false
}
],
"skills": {
"promptChars": 0,
"entries": []
},
"tools": {
"listChars": 0,
"schemaChars": 89,
"entries": [
{
"name": "session_status",
"summaryChars": 336,
"schemaChars": 89,
"propertiesCount": 2
}
]
}
},
"finalPromptText": "Reply with exactly: PONG",
"finalAssistantVisibleText": "PONG",
"finalAssistantRawText": "PONG",
"replayInvalid": false,
"livenessState": "working",
"stopReason": "stop",
"executionTrace": {
"winnerProvider": "ollama",
"winnerModel": "qwen2.5:14b",
"attempts": [
{
"provider": "ollama",
"model": "qwen2.5:14b",
"result": "success",
"stage": "assistant"
}
],
"fallbackUsed": false,
"runner": "embedded"
},
"requestShaping": {
"authMode": "auth-profile",
"thinking": "off"
},
"completion": {
"stopReason": "stop",
"finishReason": "stop"
},
"contextManagement": {
"lastTurnCompactions": 1
}
}
}
solemn ibex
after another attempt:
09:28:11 [agent/embedded] low context window: ollama/qwen2.5:14b ctx=16000 (warn<32000) source=agentContextTokens; OpenClaw is capped by agents.defaults.contextTokens, so raise that cap if you want to use more of the model context window
09:28:46 [agent/embedded] workspace bootstrap file AGENTS.md is 7809 chars (limit 400); truncating in injected context (sessionKey=agent:main:explicit:pong-test-1)
09:28:46 [agent/embedded] workspace bootstrap file SOUL.md is 1664 chars (limit 400); truncating in injected context (sessionKey=agent:main:explicit:pong-test-1)
09:28:46 [agent/embedded] remaining bootstrap budget is 2 chars (<64); skipping additional bootstrap files (sessionKey=agent:main:explicit:pong-test-1)
09:43:46 [agent/embedded] embedded run timeout: runId=pong-test-1 sessionId=pong-test-1 timeoutMs=900000
09:43:46 [agent/embedded] Profile ollama:default timed out. Trying next account...
09:43:46 [agent/embedded] embedded run failover decision: runId=pong-test-1 stage=assistant decision=surface_error reason=timeout from=ollama/qwen2.5:14b profile=sha256:9c018ec112cf