Spaces:
Runtime error
Runtime error
remove gradio
Browse files- server/app.py +24 -212
- server/models.py +30 -51
server/app.py
CHANGED
|
@@ -1,249 +1,61 @@
|
|
| 1 |
-
"""Full FastAPI app for the Eco-LLM Inference Routing environment."""
|
| 2 |
-
from __future__ import annotations
|
| 3 |
-
|
| 4 |
import uvicorn
|
| 5 |
-
|
| 6 |
-
from fastapi import FastAPI, HTTPException
|
| 7 |
-
from gradio.routes import mount_gradio_app
|
| 8 |
from openenv.core.env_server.http_server import create_app
|
| 9 |
from pydantic import BaseModel
|
| 10 |
from typing import List
|
| 11 |
|
| 12 |
-
# Your local logic imports
|
| 13 |
from server.env import EcoLLMInferenceRoutingEnvironment
|
| 14 |
from server.models import RLAction, RLObservation, Strategy, ModelChoice
|
| 15 |
from server.tasks import TASKS
|
| 16 |
|
| 17 |
-
#
|
| 18 |
-
|
| 19 |
-
|
| 20 |
-
|
| 21 |
-
|
| 22 |
-
|
| 23 |
-
|
| 24 |
-
|
|
|
|
|
|
|
| 25 |
_GRADER_BOUNDS = {
|
| 26 |
-
# task_1: 1 query, min_model=MEDIUM
|
| 27 |
-
# WORST: NONE+SMALL → wrong → -(0.8×0.1 + 0.1×1.0) = -0.18
|
| 28 |
-
# BEST: EARLY_EXIT+MEDIUM → 1 - 0.8×0.3 - 0.1×2.0 + 0.1 = 0.66
|
| 29 |
"task_1": {"worst": -0.18, "best": 0.66},
|
| 30 |
-
# task_2: 3 queries (min: SMALL, MEDIUM, LARGE); large_model_penalty=-0.2
|
| 31 |
-
# WORST: NONE+SMALL → q1 correct (SMALL ok), q2/q3 wrong
|
| 32 |
-
# = (1-0.08-0.1) + (0-0.06-0.1) + (0-0.04-0.1) = 0.52
|
| 33 |
-
# BEST: EARLY_EXIT per query (SMALL/MEDIUM/LARGE) = 0.92+0.72+0.16 = 1.80
|
| 34 |
"task_2": {"worst": 0.52, "best": 1.80},
|
| 35 |
-
# task_3: 5 queries (min: SMALL/kb, MEDIUM, SMALL/kb, LARGE, MEDIUM)
|
| 36 |
-
# WORST: NONE+SMALL → q1/q3 correct, q2/q4/q5 wrong = 1.24
|
| 37 |
-
# BEST: CALL_KB(q1)+EE+MEDIUM(q2)+USE_CACHE(q3)+EE+LARGE(q4)+EE+MEDIUM(q5)
|
| 38 |
-
# = 0.81+0.72+1.50+0.42+0.75 = 4.20
|
| 39 |
"task_3": {"worst": 1.24, "best": 4.20},
|
| 40 |
}
|
| 41 |
-
_GRADER_EPS = 0.01
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
def _run_episode(task_id: str, actions: list[dict]) -> float:
|
| 45 |
-
"""Replay a fixed list of actions and return the normalised score."""
|
| 46 |
-
env = EcoLLMInferenceRoutingEnvironment()
|
| 47 |
-
obs = env.reset(task_id=task_id)
|
| 48 |
-
episode_reward = 0.0
|
| 49 |
-
|
| 50 |
-
for raw_action in actions:
|
| 51 |
-
if obs.done:
|
| 52 |
-
break
|
| 53 |
-
action = RLAction(
|
| 54 |
-
strategy=Strategy(raw_action["strategy"]),
|
| 55 |
-
model_choice=ModelChoice(raw_action["model_choice"]),
|
| 56 |
-
exit_flag=bool(raw_action.get("exit_flag", False)),
|
| 57 |
-
)
|
| 58 |
-
obs = env.step(action)
|
| 59 |
-
episode_reward += obs.reward_details.total_reward
|
| 60 |
-
|
| 61 |
-
bounds = _GRADER_BOUNDS[task_id]
|
| 62 |
-
worst, best = bounds["worst"], bounds["best"]
|
| 63 |
-
normalised = (episode_reward - worst) / (best - worst)
|
| 64 |
-
score = max(_GRADER_EPS, min(1.0 - _GRADER_EPS, normalised))
|
| 65 |
-
return round(score, 6)
|
| 66 |
-
|
| 67 |
-
|
| 68 |
-
# ─────────────────────────────────────────
|
| 69 |
-
# 🏗️ REQUEST / RESPONSE MODELS
|
| 70 |
-
# ─────────────────────────────────────────
|
| 71 |
|
| 72 |
class GraderAction(BaseModel):
|
| 73 |
strategy: str
|
| 74 |
model_choice: str
|
| 75 |
exit_flag: bool = False
|
| 76 |
|
| 77 |
-
|
| 78 |
class GraderRequest(BaseModel):
|
| 79 |
task_id: str
|
| 80 |
actions: List[GraderAction]
|
| 81 |
|
| 82 |
-
|
| 83 |
-
class GraderResponse(BaseModel):
|
| 84 |
-
task_id: str
|
| 85 |
-
score: float # strictly in (0, 1)
|
| 86 |
-
episode_reward: float
|
| 87 |
-
worst_bound: float
|
| 88 |
-
best_bound: float
|
| 89 |
-
|
| 90 |
-
|
| 91 |
-
# 1. Create the Master App FIRST
|
| 92 |
-
app = FastAPI()
|
| 93 |
-
|
| 94 |
-
# ✅ REQUIRED FOR EVALUATOR
|
| 95 |
@app.get("/health")
|
| 96 |
def health():
|
| 97 |
return {"status": "ok"}
|
| 98 |
|
| 99 |
-
|
| 100 |
-
# ─────────────────────────────────────────
|
| 101 |
-
# 📋 GET /tasks – task catalogue
|
| 102 |
-
# ─────────────────────────────────────────
|
| 103 |
-
|
| 104 |
@app.get("/tasks")
|
| 105 |
def list_tasks():
|
| 106 |
-
""
|
| 107 |
-
return {
|
| 108 |
-
task_id: {
|
| 109 |
-
"task_id": spec.task_id,
|
| 110 |
-
"difficulty": spec.difficulty,
|
| 111 |
-
"description": spec.description,
|
| 112 |
-
"num_queries": len(spec.queries),
|
| 113 |
-
"grader_bounds": _GRADER_BOUNDS[task_id],
|
| 114 |
-
}
|
| 115 |
-
for task_id, spec in TASKS.items()
|
| 116 |
-
}
|
| 117 |
-
|
| 118 |
|
| 119 |
-
|
| 120 |
-
# 🏆 POST /grader – deterministic grader
|
| 121 |
-
# ─────────────────────────────────────────
|
| 122 |
-
|
| 123 |
-
@app.post("/grader", response_model=GraderResponse)
|
| 124 |
def grade_episode(request: GraderRequest):
|
| 125 |
-
"""
|
| 126 |
-
Deterministically grade a completed episode.
|
| 127 |
-
|
| 128 |
-
Supply the task_id and the ordered list of actions taken during the
|
| 129 |
-
episode. The endpoint replays those actions in a fresh environment
|
| 130 |
-
instance and returns a score strictly between 0 and 1.
|
| 131 |
-
|
| 132 |
-
Score formula
|
| 133 |
-
-------------
|
| 134 |
-
score = soft_clamp(
|
| 135 |
-
(episode_reward - WORST) / (BEST - WORST),
|
| 136 |
-
lo=0.01, hi=0.99
|
| 137 |
-
)
|
| 138 |
-
|
| 139 |
-
where WORST and BEST are the empirically determined lower and upper
|
| 140 |
-
bounds for each task (see GET /tasks for exact values).
|
| 141 |
-
"""
|
| 142 |
-
task_id = request.task_id
|
| 143 |
-
if task_id not in _GRADER_BOUNDS:
|
| 144 |
-
raise HTTPException(
|
| 145 |
-
status_code=400,
|
| 146 |
-
detail=f"Unknown task_id {task_id!r}. Valid: {list(_GRADER_BOUNDS)}"
|
| 147 |
-
)
|
| 148 |
-
|
| 149 |
-
raw_actions = [a.model_dump() for a in request.actions]
|
| 150 |
-
|
| 151 |
-
# Replay to measure episode_reward independently for the response
|
| 152 |
env = EcoLLMInferenceRoutingEnvironment()
|
| 153 |
-
obs = env.reset(task_id=task_id)
|
| 154 |
-
|
| 155 |
-
for
|
| 156 |
-
if obs.done:
|
| 157 |
-
|
| 158 |
-
action = RLAction(
|
| 159 |
-
strategy=Strategy(raw_action["strategy"]),
|
| 160 |
-
model_choice=ModelChoice(raw_action["model_choice"]),
|
| 161 |
-
exit_flag=bool(raw_action.get("exit_flag", False)),
|
| 162 |
-
)
|
| 163 |
obs = env.step(action)
|
| 164 |
-
|
| 165 |
-
|
| 166 |
-
bounds = _GRADER_BOUNDS[task_id]
|
| 167 |
-
worst, best = bounds["worst"], bounds["best"]
|
| 168 |
-
normalised = (episode_reward - worst) / (best - worst)
|
| 169 |
-
score = max(_GRADER_EPS, min(1.0 - _GRADER_EPS, normalised))
|
| 170 |
-
|
| 171 |
-
return GraderResponse(
|
| 172 |
-
task_id=task_id,
|
| 173 |
-
score=round(score, 6),
|
| 174 |
-
episode_reward=round(episode_reward, 6),
|
| 175 |
-
worst_bound=worst,
|
| 176 |
-
best_bound=best,
|
| 177 |
-
)
|
| 178 |
-
|
| 179 |
-
# ─────────────────────────────────────────
|
| 180 |
-
# 🎯 GRADIO UI DEFINITION
|
| 181 |
-
# ─────────────────────────────────────────
|
| 182 |
-
|
| 183 |
-
def reset_demo(task_id: str, state: dict) -> tuple:
|
| 184 |
-
demo_env = EcoLLMInferenceRoutingEnvironment()
|
| 185 |
-
obs = demo_env.reset(task_id=task_id)
|
| 186 |
-
updated_state = {"env": demo_env, "obs": obs, "history": [], "total_reward": 0.0}
|
| 187 |
-
|
| 188 |
-
log = f"🔄 Reset — Task: {task_id}\n📋 Query: {obs.query}\n🌍 Carbon: {obs.carbon_intensity:.2f}"
|
| 189 |
-
return log, "", 0.0, updated_state
|
| 190 |
-
|
| 191 |
-
def step_demo(strategy: str, model: str, exit_flag: bool, state: dict) -> tuple:
|
| 192 |
-
demo_env = state.get("env")
|
| 193 |
-
if not demo_env: return "⚠️ Reset first!", "", 0.0, state
|
| 194 |
-
|
| 195 |
-
action = RLAction(strategy=Strategy(strategy), model_choice=ModelChoice(model), exit_flag=exit_flag)
|
| 196 |
-
new_obs = demo_env.step(action)
|
| 197 |
-
rd = new_obs.reward_details
|
| 198 |
-
total_reward = state.get("total_reward", 0.0) + rd.total_reward
|
| 199 |
-
|
| 200 |
-
log = f"Step Result: {'✅' if rd.correct else '❌'}\nReward: {rd.total_reward:+.3f}"
|
| 201 |
-
return log, str(rd), round(total_reward, 3), {**state, "obs": new_obs, "total_reward": total_reward}
|
| 202 |
-
|
| 203 |
-
with gr.Blocks(title="Eco-LLM Routing") as demo:
|
| 204 |
-
demo_state = gr.State({"env": None, "obs": None, "history": [], "total_reward": 0.0})
|
| 205 |
-
gr.Markdown("# 🌱 Eco-LLM Inference Routing")
|
| 206 |
-
|
| 207 |
-
with gr.Row():
|
| 208 |
-
task_selector = gr.Dropdown(choices=["task_1", "task_2", "task_3"], value="task_1", label="Task")
|
| 209 |
-
reset_btn = gr.Button("🔄 Reset", variant="primary")
|
| 210 |
-
|
| 211 |
-
obs_box = gr.Textbox(label="State", lines=4, interactive=False)
|
| 212 |
|
| 213 |
-
|
| 214 |
-
|
| 215 |
-
|
| 216 |
-
exit_cb = gr.Checkbox(label="Exit Flag", value=False)
|
| 217 |
-
|
| 218 |
-
step_btn = gr.Button("▶️ Step")
|
| 219 |
-
reward_box = gr.Textbox(label="Step Detail", interactive=False)
|
| 220 |
-
total_num = gr.Number(label="Total Reward", interactive=False)
|
| 221 |
-
|
| 222 |
-
reset_btn.click(reset_demo, [task_selector, demo_state], [obs_box, reward_box, total_num, demo_state])
|
| 223 |
-
step_btn.click(step_demo, [strategy_dd, model_dd, exit_cb, demo_state], [obs_box, reward_box, total_num, demo_state])
|
| 224 |
-
|
| 225 |
-
# ─────────────────────────────────────────
|
| 226 |
-
# 🚀 ROUTING SETUP
|
| 227 |
-
# ─────────────────────────────────────────
|
| 228 |
-
|
| 229 |
-
# 2. MOUNT GRADIO TO THE ROOT ("/")
|
| 230 |
-
# This ensures Gradio's internal JS/CSS assets load without path issues.
|
| 231 |
-
app = mount_gradio_app(app, demo, path="/")
|
| 232 |
-
|
| 233 |
-
# 3. Create and mount OpenEnv to "/api"
|
| 234 |
-
# This keeps the environment logic accessible but gets its UI out of the way.
|
| 235 |
-
openenv_logic = create_app(
|
| 236 |
-
EcoLLMInferenceRoutingEnvironment,
|
| 237 |
-
RLAction,
|
| 238 |
-
RLObservation,
|
| 239 |
-
env_name="eco_llm_inference_routing",
|
| 240 |
-
max_concurrent_envs=4,
|
| 241 |
-
)
|
| 242 |
-
app.mount("/api", openenv_logic)
|
| 243 |
-
|
| 244 |
-
|
| 245 |
-
def main(host: str = "0.0.0.0", port: int = 7860) -> None:
|
| 246 |
-
uvicorn.run(app, host=host, port=port)
|
| 247 |
|
| 248 |
if __name__ == "__main__":
|
| 249 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
import uvicorn
|
| 2 |
+
from fastapi import HTTPException
|
|
|
|
|
|
|
| 3 |
from openenv.core.env_server.http_server import create_app
|
| 4 |
from pydantic import BaseModel
|
| 5 |
from typing import List
|
| 6 |
|
|
|
|
| 7 |
from server.env import EcoLLMInferenceRoutingEnvironment
|
| 8 |
from server.models import RLAction, RLObservation, Strategy, ModelChoice
|
| 9 |
from server.tasks import TASKS
|
| 10 |
|
| 11 |
+
# 1. Create native OpenEnv app
|
| 12 |
+
app = create_app(
|
| 13 |
+
EcoLLMInferenceRoutingEnvironment,
|
| 14 |
+
RLAction,
|
| 15 |
+
RLObservation,
|
| 16 |
+
env_name="eco_llm_inference_routing",
|
| 17 |
+
max_concurrent_envs=4,
|
| 18 |
+
)
|
| 19 |
+
|
| 20 |
+
# 2. Grading Metadata
|
| 21 |
_GRADER_BOUNDS = {
|
|
|
|
|
|
|
|
|
|
| 22 |
"task_1": {"worst": -0.18, "best": 0.66},
|
|
|
|
|
|
|
|
|
|
|
|
|
| 23 |
"task_2": {"worst": 0.52, "best": 1.80},
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
"task_3": {"worst": 1.24, "best": 4.20},
|
| 25 |
}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 26 |
|
| 27 |
class GraderAction(BaseModel):
|
| 28 |
strategy: str
|
| 29 |
model_choice: str
|
| 30 |
exit_flag: bool = False
|
| 31 |
|
|
|
|
| 32 |
class GraderRequest(BaseModel):
|
| 33 |
task_id: str
|
| 34 |
actions: List[GraderAction]
|
| 35 |
|
| 36 |
+
# 3. Add required extra endpoints to the OpenEnv app
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 37 |
@app.get("/health")
|
| 38 |
def health():
|
| 39 |
return {"status": "ok"}
|
| 40 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 41 |
@app.get("/tasks")
|
| 42 |
def list_tasks():
|
| 43 |
+
return {tid: {"task_id": tid, "bounds": _GRADER_BOUNDS.get(tid)} for tid in TASKS.keys()}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 44 |
|
| 45 |
+
@app.post("/grader")
|
|
|
|
|
|
|
|
|
|
|
|
|
| 46 |
def grade_episode(request: GraderRequest):
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 47 |
env = EcoLLMInferenceRoutingEnvironment()
|
| 48 |
+
obs = env.reset(task_id=request.task_id)
|
| 49 |
+
total_reward = 0.0
|
| 50 |
+
for a in request.actions:
|
| 51 |
+
if obs.done: break
|
| 52 |
+
action = RLAction(strategy=Strategy(a.strategy), model_choice=ModelChoice(a.model_choice), exit_flag=a.exit_flag)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 53 |
obs = env.step(action)
|
| 54 |
+
total_reward += obs.reward_details.total_reward
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 55 |
|
| 56 |
+
bounds = _GRADER_BOUNDS[request.task_id]
|
| 57 |
+
score = (total_reward - bounds["worst"]) / (bounds["best"] - bounds["worst"])
|
| 58 |
+
return {"task_id": request.task_id, "score": max(0.01, min(0.99, score)), "reward": total_reward}
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 59 |
|
| 60 |
if __name__ == "__main__":
|
| 61 |
+
uvicorn.run(app, host="0.0.0.0", port=7860)
|
server/models.py
CHANGED
|
@@ -1,14 +1,10 @@
|
|
| 1 |
"""Pydantic models for the Eco-LLM Inference Routing environment."""
|
| 2 |
-
|
| 3 |
from __future__ import annotations
|
| 4 |
-
|
| 5 |
from enum import Enum
|
| 6 |
from typing import Any
|
| 7 |
-
|
| 8 |
from openenv.core.env_server.types import Action, Observation, State
|
| 9 |
from pydantic import BaseModel, Field
|
| 10 |
|
| 11 |
-
|
| 12 |
class Strategy(str, Enum):
|
| 13 |
NONE = "NONE"
|
| 14 |
USE_CACHE = "USE_CACHE"
|
|
@@ -17,66 +13,49 @@ class Strategy(str, Enum):
|
|
| 17 |
WAIT = "WAIT"
|
| 18 |
CALL_KB = "CALL_KB"
|
| 19 |
|
| 20 |
-
|
| 21 |
class ModelChoice(str, Enum):
|
| 22 |
SMALL = "SMALL"
|
| 23 |
MEDIUM = "MEDIUM"
|
| 24 |
LARGE = "LARGE"
|
| 25 |
|
| 26 |
-
|
| 27 |
class RLAction(Action):
|
| 28 |
"""Action taken by the routing agent at each step."""
|
| 29 |
-
|
| 30 |
-
strategy: Strategy = Field(.
|
| 31 |
-
model_choice: ModelChoice = Field(.
|
| 32 |
-
exit_flag: bool = Field(
|
| 33 |
-
|
| 34 |
|
| 35 |
class RLReward(BaseModel):
|
| 36 |
-
|
| 37 |
-
|
| 38 |
-
|
| 39 |
-
|
| 40 |
-
|
| 41 |
-
|
| 42 |
-
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
correct: bool = Field(..., description="Whether the selected answer was correct")
|
| 46 |
-
strategy_trace: list[str] = Field(
|
| 47 |
-
default_factory=list,
|
| 48 |
-
description="Ordered record of sub-actions executed during the step",
|
| 49 |
-
)
|
| 50 |
-
|
| 51 |
|
| 52 |
class RLObservation(Observation):
|
| 53 |
-
|
| 54 |
-
|
| 55 |
-
|
| 56 |
-
|
| 57 |
-
|
| 58 |
-
|
| 59 |
-
|
| 60 |
-
|
| 61 |
-
reward_details: RLReward = Field(..., description="Structured reward payload")
|
| 62 |
-
kb_available: bool = Field(default=False, description="Whether KB is available for current query")
|
| 63 |
-
reward: float = Field(default=0.0, description="Step reward for this transition")
|
| 64 |
-
done: bool = Field(default=False, description="Whether episode has terminated")
|
| 65 |
-
metadata: dict[str, Any] = Field(default_factory=dict, description="Additional state metadata")
|
| 66 |
-
|
| 67 |
|
| 68 |
class RLState(State):
|
| 69 |
-
|
| 70 |
-
|
| 71 |
-
|
| 72 |
-
|
| 73 |
-
|
| 74 |
-
|
| 75 |
-
|
| 76 |
-
|
| 77 |
-
|
| 78 |
-
cache_contents: list[str] = Field(default_factory=list, description="Current cache contents")
|
| 79 |
-
carbon_history_index: int = Field(default=0, ge=0, description="Current carbon history pointer")
|
| 80 |
|
| 81 |
def public_payload(self) -> dict[str, Any]:
|
| 82 |
return self.model_dump()
|
|
|
|
| 1 |
"""Pydantic models for the Eco-LLM Inference Routing environment."""
|
|
|
|
| 2 |
from __future__ import annotations
|
|
|
|
| 3 |
from enum import Enum
|
| 4 |
from typing import Any
|
|
|
|
| 5 |
from openenv.core.env_server.types import Action, Observation, State
|
| 6 |
from pydantic import BaseModel, Field
|
| 7 |
|
|
|
|
| 8 |
class Strategy(str, Enum):
|
| 9 |
NONE = "NONE"
|
| 10 |
USE_CACHE = "USE_CACHE"
|
|
|
|
| 13 |
WAIT = "WAIT"
|
| 14 |
CALL_KB = "CALL_KB"
|
| 15 |
|
|
|
|
| 16 |
class ModelChoice(str, Enum):
|
| 17 |
SMALL = "SMALL"
|
| 18 |
MEDIUM = "MEDIUM"
|
| 19 |
LARGE = "LARGE"
|
| 20 |
|
|
|
|
| 21 |
class RLAction(Action):
|
| 22 |
"""Action taken by the routing agent at each step."""
|
| 23 |
+
# We use explicit defaults to ensure the UI renders dropdowns
|
| 24 |
+
strategy: Strategy = Field(default=Strategy.NONE, description="Routing strategy to apply")
|
| 25 |
+
model_choice: ModelChoice = Field(default=ModelChoice.SMALL, description="Primary model size")
|
| 26 |
+
exit_flag: bool = Field(default=False, description="Terminate cascade early")
|
|
|
|
| 27 |
|
| 28 |
class RLReward(BaseModel):
|
| 29 |
+
score: float = 0.0
|
| 30 |
+
energy_cost: float = 0.0
|
| 31 |
+
latency_cost: float = 0.0
|
| 32 |
+
carbon_penalty: float = 0.0
|
| 33 |
+
latency_penalty: float = 0.0
|
| 34 |
+
bonuses: float = 0.0
|
| 35 |
+
total_reward: float = 0.0
|
| 36 |
+
correct: bool = False
|
| 37 |
+
strategy_trace: list[str] = Field(default_factory=list)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 38 |
|
| 39 |
class RLObservation(Observation):
|
| 40 |
+
query: str = ""
|
| 41 |
+
cache_contents: list[str] = Field(default_factory=list)
|
| 42 |
+
carbon_intensity: float = 0.0
|
| 43 |
+
reward_details: RLReward = Field(default_factory=RLReward)
|
| 44 |
+
kb_available: bool = False
|
| 45 |
+
reward: float = 0.0
|
| 46 |
+
done: bool = False
|
| 47 |
+
metadata: dict[str, Any] = Field(default_factory=dict)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 48 |
|
| 49 |
class RLState(State):
|
| 50 |
+
episode_id: str = ""
|
| 51 |
+
step_count: int = 0
|
| 52 |
+
task_id: str = ""
|
| 53 |
+
step_index: int = 0
|
| 54 |
+
query_index: int = 0
|
| 55 |
+
total_energy: float = 0.0
|
| 56 |
+
total_latency: float = 0.0
|
| 57 |
+
cache_contents: list[str] = Field(default_factory=list)
|
| 58 |
+
carbon_history_index: int = 0
|
|
|
|
|
|
|
| 59 |
|
| 60 |
def public_payload(self) -> dict[str, Any]:
|
| 61 |
return self.model_dump()
|