Spaces:
Running
Running
Implement tests for inference helper functions
Browse filesAdd unit tests for inference helpers including action parsing and fallback policies.
- tests/test_inference.py +181 -0
tests/test_inference.py
ADDED
|
@@ -0,0 +1,181 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"""
|
| 2 |
+
Test inference helpers.
|
| 3 |
+
|
| 4 |
+
These tests focus on the local Ollama/DeepSeek path where the model often
|
| 5 |
+
wraps JSON in reasoning text or code fences.
|
| 6 |
+
"""
|
| 7 |
+
|
| 8 |
+
import sys
|
| 9 |
+
import os
|
| 10 |
+
|
| 11 |
+
sys.path.insert(0, os.path.join(os.path.dirname(__file__), ".."))
|
| 12 |
+
|
| 13 |
+
from inference import (
|
| 14 |
+
_is_ollama_backend,
|
| 15 |
+
_should_replace_with_fallback,
|
| 16 |
+
clamp_submission_score,
|
| 17 |
+
get_fallback_action,
|
| 18 |
+
parse_action,
|
| 19 |
+
)
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
class TestParseAction:
|
| 23 |
+
def test_parses_reasoning_plus_fenced_json(self):
|
| 24 |
+
response = """
|
| 25 |
+
<think>
|
| 26 |
+
I should bring up the hydro dam first.
|
| 27 |
+
</think>
|
| 28 |
+
```json
|
| 29 |
+
{"action_type":"dispatch_generation","node_id":"NODE_01","mw":600}
|
| 30 |
+
```
|
| 31 |
+
"""
|
| 32 |
+
action = parse_action(response)
|
| 33 |
+
assert action == {
|
| 34 |
+
"action_type": "dispatch_generation",
|
| 35 |
+
"node_id": "NODE_01",
|
| 36 |
+
"mw": 600,
|
| 37 |
+
}
|
| 38 |
+
|
| 39 |
+
def test_parses_single_quoted_payload(self):
|
| 40 |
+
response = "{'action_type': 'toggle_circuit_breaker', 'edge_id': 'LINE_28', 'status': 'closed'}"
|
| 41 |
+
action = parse_action(response)
|
| 42 |
+
assert action == {
|
| 43 |
+
"action_type": "toggle_circuit_breaker",
|
| 44 |
+
"edge_id": "LINE_28",
|
| 45 |
+
"status": "CLOSED",
|
| 46 |
+
}
|
| 47 |
+
|
| 48 |
+
def test_parses_action_wrapped_in_list(self):
|
| 49 |
+
response = """
|
| 50 |
+
Here is the action:
|
| 51 |
+
[
|
| 52 |
+
{"action_type":"advance_tick"}
|
| 53 |
+
]
|
| 54 |
+
"""
|
| 55 |
+
action = parse_action(response)
|
| 56 |
+
assert action == {"action_type": "advance_tick"}
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
class TestClampSubmissionScore:
|
| 60 |
+
def test_clamp_keeps_scores_inside_open_interval(self):
|
| 61 |
+
assert 0.0 < clamp_submission_score(0.0) < 1.0
|
| 62 |
+
assert 0.0 < clamp_submission_score(1.0) < 1.0
|
| 63 |
+
assert clamp_submission_score(0.42) == 0.42
|
| 64 |
+
|
| 65 |
+
|
| 66 |
+
class TestOllamaDetection:
|
| 67 |
+
def test_local_ollama_url_detected_automatically(self):
|
| 68 |
+
assert _is_ollama_backend("http://localhost:11434/v1") is True
|
| 69 |
+
|
| 70 |
+
def test_remote_openai_compatible_url_requires_override(self):
|
| 71 |
+
assert _is_ollama_backend("https://example.com/v1") is False
|
| 72 |
+
assert _is_ollama_backend("https://example.com/v1", compat_override="1") is True
|
| 73 |
+
|
| 74 |
+
|
| 75 |
+
class TestFallbackPolicy:
|
| 76 |
+
def test_cascade_overload_replaces_non_isolation_action_while_line_is_overloaded(self):
|
| 77 |
+
obs = {
|
| 78 |
+
"topology_graph": {
|
| 79 |
+
"nodes": [
|
| 80 |
+
{"id": "NODE_01", "energized": True, "capacity_mw": 1500},
|
| 81 |
+
],
|
| 82 |
+
"edges": [
|
| 83 |
+
{
|
| 84 |
+
"id": "LINE_29",
|
| 85 |
+
"status": "LIVE",
|
| 86 |
+
"current_load_mw": 950.0,
|
| 87 |
+
"capacity_mw": 1000.0,
|
| 88 |
+
},
|
| 89 |
+
],
|
| 90 |
+
},
|
| 91 |
+
"telemetry_stream": [[
|
| 92 |
+
{"node_id": "NODE_01", "generation_mw": 600.0},
|
| 93 |
+
]],
|
| 94 |
+
}
|
| 95 |
+
|
| 96 |
+
action = {"action_type": "dispatch_generation", "node_id": "NODE_01", "mw": 100}
|
| 97 |
+
assert _should_replace_with_fallback(task_id=2, tick=0, action_dict=action, obs_dict=obs)
|
| 98 |
+
|
| 99 |
+
def test_task3_replaces_early_dispatch_before_logs(self):
|
| 100 |
+
obs = {
|
| 101 |
+
"network_packet_logs": [],
|
| 102 |
+
"metadata": {"active_spoofs": ["NODE_14"]},
|
| 103 |
+
}
|
| 104 |
+
action = {"action_type": "dispatch_generation", "node_id": "NODE_09", "mw": 100}
|
| 105 |
+
assert _should_replace_with_fallback(task_id=3, tick=0, action_dict=action, obs_dict=obs)
|
| 106 |
+
|
| 107 |
+
def test_task3_replaces_early_quarantine_before_estimation(self):
|
| 108 |
+
obs = {
|
| 109 |
+
"network_packet_logs": [{"source_node": "NODE_14", "anomaly_flag": True}],
|
| 110 |
+
"metadata": {"active_spoofs": ["NODE_14"]},
|
| 111 |
+
}
|
| 112 |
+
action = {"action_type": "quarantine_scada_node", "node_id": "NODE_14"}
|
| 113 |
+
assert _should_replace_with_fallback(task_id=3, tick=1, action_dict=action, obs_dict=obs)
|
| 114 |
+
|
| 115 |
+
def test_task3_replaces_dispatch_before_quarantine(self):
|
| 116 |
+
obs = {
|
| 117 |
+
"network_packet_logs": [{"source_node": "NODE_14", "anomaly_flag": True}],
|
| 118 |
+
"last_state_estimation": {"consistent": False, "violation_node": "NODE_14"},
|
| 119 |
+
"metadata": {"active_spoofs": ["NODE_14"]},
|
| 120 |
+
}
|
| 121 |
+
action = {"action_type": "dispatch_generation", "node_id": "NODE_09", "mw": 100}
|
| 122 |
+
assert _should_replace_with_fallback(task_id=3, tick=2, action_dict=action, obs_dict=obs)
|
| 123 |
+
|
| 124 |
+
def test_task3_allows_reroute_after_quarantine(self):
|
| 125 |
+
obs = {
|
| 126 |
+
"topology_graph": {
|
| 127 |
+
"nodes": [
|
| 128 |
+
{"id": "NODE_09", "energized": True, "capacity_mw": 900},
|
| 129 |
+
],
|
| 130 |
+
"edges": [],
|
| 131 |
+
},
|
| 132 |
+
"telemetry_stream": [[
|
| 133 |
+
{"node_id": "NODE_09", "generation_mw": 400.0},
|
| 134 |
+
]],
|
| 135 |
+
"network_packet_logs": [{"source_node": "NODE_14", "anomaly_flag": True}],
|
| 136 |
+
"last_state_estimation": {"consistent": False, "violation_node": "NODE_14"},
|
| 137 |
+
"metadata": {"active_spoofs": []},
|
| 138 |
+
}
|
| 139 |
+
action = {"action_type": "dispatch_generation", "node_id": "NODE_09", "mw": 100}
|
| 140 |
+
assert _should_replace_with_fallback(task_id=3, tick=3, action_dict=action, obs_dict=obs) is False
|
| 141 |
+
|
| 142 |
+
def test_task4_replaces_non_injection_opening_move(self):
|
| 143 |
+
obs = {
|
| 144 |
+
"topology_graph": {"nodes": [], "edges": []},
|
| 145 |
+
"telemetry_stream": [],
|
| 146 |
+
}
|
| 147 |
+
action = {"action_type": "run_state_estimation", "subgraph": ["NODE_17", "NODE_20"]}
|
| 148 |
+
assert _should_replace_with_fallback(task_id=4, tick=0, action_dict=action, obs_dict=obs)
|
| 149 |
+
|
| 150 |
+
def test_task3_replaces_redundant_estimation_after_quarantine(self):
|
| 151 |
+
obs = {
|
| 152 |
+
"network_packet_logs": [{"source_node": "NODE_14", "anomaly_flag": True}],
|
| 153 |
+
"last_state_estimation": {"consistent": False, "violation_node": "NODE_14"},
|
| 154 |
+
"metadata": {"active_spoofs": []},
|
| 155 |
+
}
|
| 156 |
+
action = {"action_type": "run_state_estimation", "subgraph": ["NODE_14", "NODE_15"]}
|
| 157 |
+
assert _should_replace_with_fallback(task_id=3, tick=4, action_dict=action, obs_dict=obs)
|
| 158 |
+
|
| 159 |
+
def test_black_start_fallback_only_makes_minimal_progress(self):
|
| 160 |
+
obs = {
|
| 161 |
+
"topology_graph": {
|
| 162 |
+
"nodes": [
|
| 163 |
+
{"id": "NODE_01", "energized": True},
|
| 164 |
+
{"id": "NODE_03", "energized": False},
|
| 165 |
+
],
|
| 166 |
+
"edges": [
|
| 167 |
+
{"id": "LINE_02", "status": "TRIPPED"},
|
| 168 |
+
],
|
| 169 |
+
},
|
| 170 |
+
"telemetry_stream": [[
|
| 171 |
+
{"node_id": "NODE_01", "generation_mw": 600.0},
|
| 172 |
+
{"node_id": "NODE_03", "generation_mw": 0.0},
|
| 173 |
+
]],
|
| 174 |
+
}
|
| 175 |
+
|
| 176 |
+
action = get_fallback_action(task_id=5, tick=3, obs_dict=obs)
|
| 177 |
+
assert action == {
|
| 178 |
+
"action_type": "toggle_circuit_breaker",
|
| 179 |
+
"edge_id": "LINE_02",
|
| 180 |
+
"status": "CLOSED",
|
| 181 |
+
}
|