Repository navigation
Expand file tree
/
Copy pathsample_trajectory.py
More file actions
101 lines (86 loc) · 3.25 KB
/
Copy pathsample_trajectory.py
File metadata and controls
101 lines (86 loc) · 3.25 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
"""Runnable CPU contract example using a toy environment and backend.
This is an integration example, not an agent-quality or GPU benchmark.
Replace the environment/backend with real implementations of the same methods.
"""
from cachebridge.trajectory import (
Generation,
Message,
Observation,
RouteDecision,
TaskRef,
)
from cachebridge.agents import CacheBridgePolicy, sample_trajectory
from cachebridge.offline.trajectory import handoff_samples
class ExampleEnvironment:
def reset(self, task, *, seed):
self.messages = (Message("user", "Look up the order and resolve the request."),)
self.calls = 0
return Observation(self.messages)
def step(self, response):
self.calls += 1
self.messages += (response,)
if self.calls == 3:
return Observation(
self.messages, done=True, stop_reason="environment_terminal"
)
self.messages += (Message("tool", "Lookup completed.", name="lookup_order"),)
return Observation(self.messages)
class ExampleRouter:
def choose(self, context):
ready = any(
m.role == "tool" and m.name == "lookup_order" for m in context.messages
)
return RouteDecision(
"large" if ready else "small", "resolve_after_lookup" if ready else "intake"
)
class ExampleBackend:
def begin_episode(self, trajectory_id):
self.calls = 0
def generate(self, request):
prompts = ((1, 2), (1, 2, 3, 4), (1, 2, 3, 4, 5, 6, 7))
outputs = ((3,), (5, 6), (8, 9))
step = self.calls
self.calls += 1
mode = request.cache_plan.mode
return Generation(
response=Message("assistant", ("look up", "prepare", "resolve")[step]),
prompt_ids=prompts[step],
completion_ids=outputs[step],
tokenizer_id="toy-tokenizer-v1",
model_id=request.model_id,
executed_mode=mode,
reused_tokens=3 if mode == "cachebridge" else (6 if mode == "reuse" else 0),
handoff_prefix_ids=(1, 2, 3) if step == 1 else (),
mapper_id=request.cache_plan.mapper_id if mode == "cachebridge" else None,
)
def end_episode(self, trajectory_id):
pass
def main():
task = TaskRef("example", "orders", "train", "0")
for enabled in (False, True):
trajectory = sample_trajectory(
task,
run_id="example/mapped" if enabled else "example/native",
environment=ExampleEnvironment(),
router=ExampleRouter(),
backend=ExampleBackend(),
models=("small", "large"),
cache_policy=CacheBridgePolicy(
{("small", "large"): "local-mapper-v1"}, enabled=enabled
),
seed=7,
)
assert trajectory.status == "terminal", trajectory.error
print(
"CacheBridge:",
enabled,
"models:",
[t.route.model_id for t in trajectory.turns],
"execution:",
[t.generation.executed_mode for t in trajectory.turns],
)
if not enabled:
samples = handoff_samples(trajectory)
print("Native TRAIN sample:", samples[0].to_dict())
if __name__ == "__main__":
main()