File size: 12,468 Bytes
bc1c255 52e5f26 5309389 52e5f26 5309389 52e5f26 603d970 52e5f26 5309389 52e5f26 5309389 bc1c255 52e5f26 bc1c255 52e5f26 5309389 bc1c255 52e5f26 bc1c255 5309389 bc1c255 52e5f26 bc1c255 52e5f26 bc1c255 3858b5d 9e08fe1 bc1c255 9e08fe1 bc1c255 9e08fe1 bc1c255 9e08fe1 3858b5d bc1c255 5309389 52e5f26 5309389 52e5f26 5309389 52e5f26 5309389 bc1c255 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 | """
R2-Router: LLM Router with Joint Model-Budget Optimization
Self-contained inference module. Routes queries to the optimal (model, token_budget)
pair by predicting per-query quality and cost using KNN.
Usage:
from router import R2Router
# Option A: Local vLLM (loads Qwen3-0.6B on first call)
router = R2Router.from_pretrained(path)
result = router.route_text("What is the capital of France?")
# Option B: Remote vLLM server (no local GPU needed for embedding)
# Start server: vllm serve Qwen/Qwen3-0.6B --runner pooling
router = R2Router.from_pretrained(path, embed_url="http://localhost:8000")
result = router.route_text("What is the capital of France?")
# Option C: Pre-computed embedding
result = router.route(embedding) # np.ndarray (1024,)
"""
import os
import json
import numpy as np
import joblib
from typing import Dict, List, Optional, Union
from sklearn.neighbors import KNeighborsRegressor
class R2Router:
"""
R2-Router: Routes queries to optimal (LLM, token_budget) pair.
Uses KNN to predict quality for each (model, budget) combination,
then selects the pair that maximizes:
risk = (1 - lambda) * quality - lambda * tokens * price / 1e6
"""
def __init__(
self,
quality_knns: Dict[str, Dict[str, KNeighborsRegressor]],
token_knns: Dict[str, KNeighborsRegressor],
model_prices: Dict[str, float],
model_names: Dict[str, str],
budgets: Dict[str, int],
lambda_val: float = 0.999,
embed_url: Optional[str] = None,
):
self.quality_knns = quality_knns # {model: {budget: KNN}}
self.token_knns = token_knns # {model: KNN}
self.model_prices = model_prices # {model: price_per_million_output_tokens}
self.model_names = model_names # {short_name: full_name}
self.budgets = budgets # {budget_name: token_limit}
self.lambda_val = lambda_val
self.embed_url = embed_url # vLLM server URL, e.g. "http://localhost:8000"
self._embedder = None
@classmethod
def from_pretrained(
cls,
path: str,
lambda_val: float = 0.999,
embed_url: Optional[str] = None,
) -> "R2Router":
"""
Load pre-trained KNN checkpoints.
Args:
path: Local directory or HuggingFace repo ID (e.g., "JiaqiXue/r2-router")
lambda_val: Cost-accuracy tradeoff (higher = more cost-sensitive)
embed_url: vLLM server URL for embedding (e.g., "http://localhost:8000").
If None, loads Qwen3-0.6B locally on first route_text() call.
"""
if not os.path.isdir(path):
path = cls._download_from_hf(path)
with open(os.path.join(path, "config.json")) as f:
config = json.load(f)
ckpt_dir = os.path.join(path, "checkpoints")
quality_knns = {}
token_knns = {}
for model_name in config["models"]:
quality_knns[model_name] = {}
for budget_name in config["budgets"]:
ckpt_path = os.path.join(ckpt_dir, f"quality_knn_{model_name}_{budget_name}.joblib")
if os.path.exists(ckpt_path):
quality_knns[model_name][budget_name] = joblib.load(ckpt_path)
tok_path = os.path.join(ckpt_dir, f"token_knn_{model_name}.joblib")
if os.path.exists(tok_path):
token_knns[model_name] = joblib.load(tok_path)
model_prices = {
mn: cfg["output_price_per_million"]
for mn, cfg in config["models"].items()
}
model_names = {
mn: cfg["full_name"]
for mn, cfg in config["models"].items()
}
return cls(
quality_knns=quality_knns,
token_knns=token_knns,
model_prices=model_prices,
model_names=model_names,
budgets=config["budgets"],
lambda_val=lambda_val,
embed_url=embed_url,
)
@classmethod
def from_training_data(
cls,
path: str,
k: int = 80,
lambda_val: float = 0.999,
) -> "R2Router":
"""
Train KNN from scratch using the provided training data.
Args:
path: Local directory or HuggingFace repo ID
k: Number of KNN neighbors
lambda_val: Cost-accuracy tradeoff
"""
if not os.path.isdir(path):
path = cls._download_from_hf(path)
with open(os.path.join(path, "config.json")) as f:
config = json.load(f)
X_train = np.load(os.path.join(path, "training_data", "embeddings.npy"))
with open(os.path.join(path, "training_data", "labels.json")) as f:
labels = json.load(f)
print(f"Training router on {len(X_train)} samples (k={k})...")
quality_knns = {}
token_knns = {}
n_quality = 0
n_token = 0
for model_name, model_labels in labels.items():
quality_knns[model_name] = {}
for budget_name, bdata in model_labels.items():
acc = np.array([x if x is not None else np.nan for x in bdata["accuracy"]])
valid = ~np.isnan(acc)
if valid.sum() < 3:
continue
knn = KNeighborsRegressor(
n_neighbors=min(k, int(valid.sum()) - 1),
metric="cosine",
weights="distance",
)
knn.fit(X_train[valid], acc[valid])
quality_knns[model_name][budget_name] = knn
n_quality += 1
if "concise" in model_labels and "output_tokens" in model_labels["concise"]:
tok = np.array([x if x is not None else np.nan for x in model_labels["concise"]["output_tokens"]])
valid = ~np.isnan(tok)
if valid.sum() >= 3:
tknn = KNeighborsRegressor(
n_neighbors=min(k, int(valid.sum()) - 1),
metric="cosine",
weights="distance",
)
tknn.fit(X_train[valid], tok[valid])
token_knns[model_name] = tknn
n_token += 1
print(f"Trained {n_quality} quality predictors + {n_token} token predictors for {len(quality_knns)} models.")
model_prices = {
mn: cfg["output_price_per_million"]
for mn, cfg in config["models"].items()
}
model_names = {
mn: cfg["full_name"]
for mn, cfg in config["models"].items()
}
return cls(
quality_knns=quality_knns,
token_knns=token_knns,
model_prices=model_prices,
model_names=model_names,
budgets=config["budgets"],
lambda_val=lambda_val,
)
@staticmethod
def _download_from_hf(repo_id: str) -> str:
"""Download model from Hugging Face Hub."""
try:
from huggingface_hub import snapshot_download
return snapshot_download(repo_id)
except ImportError:
raise ImportError(
"huggingface_hub is required to download from HF. "
"Install with: pip install huggingface_hub"
)
def embed(self, queries: Union[str, List[str]]) -> np.ndarray:
"""
Embed queries using Qwen3-0.6B.
If embed_url is set, uses a remote vLLM server (OpenAI-compatible API).
Otherwise, loads Qwen3-0.6B locally via vLLM (on first call).
Args:
queries: Single query string or list of queries
Returns:
numpy array of shape (N, 1024)
"""
if isinstance(queries, str):
queries = [queries]
if self.embed_url:
return self._embed_remote(queries)
return self._embed_local(queries)
def _embed_remote(self, queries: List[str]) -> np.ndarray:
"""Embed via a running vLLM server (OpenAI-compatible embeddings API)."""
import urllib.request
url = self.embed_url.rstrip("/") + "/v1/embeddings"
payload = json.dumps({
"model": "Qwen/Qwen3-0.6B",
"input": queries,
}).encode()
req = urllib.request.Request(
url, data=payload,
headers={"Content-Type": "application/json"},
)
with urllib.request.urlopen(req) as resp:
result = json.loads(resp.read())
embeddings = [item["embedding"] for item in sorted(result["data"], key=lambda x: x["index"])]
return np.array(embeddings)
def _embed_local(self, queries: List[str]) -> np.ndarray:
"""Embed by loading Qwen3-0.6B locally via vLLM."""
if self._embedder is None:
try:
from vllm import LLM
except ImportError:
raise ImportError(
"vLLM is required for local embedding. "
"Install with: uv pip install vllm\n"
"Or start a vLLM server and pass embed_url to from_pretrained()."
)
self._embedder = LLM(
model="Qwen/Qwen3-0.6B",
runner="pooling",
trust_remote_code=True,
dtype="half",
)
outputs = self._embedder.embed(queries)
return np.array([o.outputs.embedding for o in outputs])
def route_text(
self,
query: Union[str, List[str]],
lambda_val: Optional[float] = None,
) -> Union[Dict, List[Dict]]:
"""
Route text query(ies) end-to-end: embed with Qwen3-0.6B, then route.
Args:
query: Single query string or list of queries
lambda_val: Override default lambda
Returns:
Routing decision dict (single) or list of dicts (batch)
"""
embeddings = self.embed(query)
if isinstance(query, str):
return self.route(embeddings[0], lambda_val)
return self.route_batch(embeddings, lambda_val)
def route(
self,
embedding: np.ndarray,
lambda_val: Optional[float] = None,
) -> Dict:
"""
Route a query to the optimal (model, token_budget) pair.
Args:
embedding: Query embedding vector, shape (1024,) or (1, 1024)
lambda_val: Override default lambda (higher = more cost-sensitive)
Returns:
Dict with keys: model, model_full_name, budget, token_limit,
predicted_quality, predicted_cost, risk, all_options
"""
if embedding.ndim == 1:
embedding = embedding.reshape(1, -1)
lam = lambda_val if lambda_val is not None else self.lambda_val
all_options = []
for mn in self.quality_knns:
price = self.model_prices.get(mn, 0)
if mn in self.token_knns:
tok = max(1.0, float(self.token_knns[mn].predict(embedding)[0]))
else:
tok = 50.0
for budget_name, knn in self.quality_knns[mn].items():
q = float(knn.predict(embedding)[0])
risk = (1 - lam) * q - lam * tok * price / 1e6
all_options.append({
"model": mn,
"model_full_name": self.model_names.get(mn, mn),
"budget": budget_name,
"token_limit": self.budgets.get(budget_name, budget_name),
"predicted_quality": q,
"predicted_tokens": tok,
"predicted_cost": tok * price / 1e6,
"risk": risk,
})
if not all_options:
raise RuntimeError("No valid routing options")
best = max(all_options, key=lambda x: x["risk"])
best["all_options"] = all_options
return best
def route_batch(
self,
embeddings: np.ndarray,
lambda_val: Optional[float] = None,
) -> List[Dict]:
"""
Route a batch of queries.
Args:
embeddings: Query embeddings, shape (N, 1024)
lambda_val: Override default lambda
Returns:
List of routing decisions
"""
return [self.route(embeddings[i], lambda_val) for i in range(len(embeddings))]
|