File size: 16,697 Bytes
21c7db9 | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 | """LLM provider runtime with Transformers-first fallback order.
The runtime is intentionally conservative: if an LLM backend is unavailable or
errors, selection falls back to deterministic local ranking.
"""
from __future__ import annotations
from dataclasses import dataclass
import json
import os
import re
import shutil
import subprocess
import time
from pathlib import Path
from typing import Any
from app.common.types import CandidateAction
from app.models.policy.active_model import active_model_status, available_artifact_path
from app.models.policy.safety_ranker import rank_candidates
def _transformers_low_cpu_mem() -> bool:
"""Use lazy/meta init only on CUDA; on CPU it often breaks ``.to(device)`` (meta tensors)."""
try:
import torch
return torch.cuda.is_available()
except Exception:
return False
def _peft_base_model_id(artifact_path: str, status: dict[str, Any], fallback: str) -> str:
cfg = Path(artifact_path) / "adapter_config.json"
if cfg.is_file():
try:
payload = json.loads(cfg.read_text(encoding="utf-8"))
raw = payload.get("base_model_name_or_path")
if isinstance(raw, str) and raw.strip():
return raw.strip()
except Exception:
pass
return str(status.get("base_model") or fallback)
def _env_truthy(name: str, default: bool = False) -> bool:
raw = os.getenv(name)
if raw is None:
return default
return raw.strip().lower() in {"1", "true", "yes", "on"}
def default_provider_preference() -> tuple[str, ...]:
raw = os.getenv("POLYGUARD_PROVIDER_PREFERENCE", "").strip()
if raw:
order = tuple(p.strip().lower() for p in raw.split(",") if p.strip())
cleaned = tuple(p for p in order if p in {"ollama", "transformers"})
if cleaned:
return cleaned
if _env_truthy("POLYGUARD_ENABLE_OLLAMA"):
return ("ollama", "transformers")
return ("transformers",)
def _extract_candidate_id(text: str, legal_ids: set[str]) -> str | None:
lowered = text.lower()
for candidate_id in sorted(legal_ids):
if candidate_id.lower() in lowered:
return candidate_id
return None
def _compact_prompt(candidates: list[CandidateAction], prompt: dict[str, Any]) -> str:
compact_candidates = [
{
"candidate_id": c.candidate_id,
"mode": c.mode.value,
"action_type": c.action_type.value,
"target_drug": c.target_drug,
"replacement_drug": c.replacement_drug,
"dose_bucket": c.dose_bucket.value,
"safety_delta": c.estimated_safety_delta,
"uncertainty": c.uncertainty_score,
"legal": c.legality_precheck,
"tags": c.rationale_tags[:4],
}
for c in candidates
]
payload = {
"instruction": "Select the safest legal medication action candidate_id.",
"context": prompt,
"candidate_ids": [c.candidate_id for c in candidates],
"candidates": compact_candidates,
"answer": "",
"format": "Return candidate_id=<one candidate_id>; rationale=<brief clinical reason>.",
}
return json.dumps(payload, ensure_ascii=True)
@dataclass(slots=True)
class ProviderSelection:
provider: str
candidate_id: str
rationale: str
latency_ms: float
raw_output: str = ""
class OllamaProvider:
name = "ollama"
def __init__(self, model_name: str) -> None:
self.model_name = model_name
self._last_error = ""
def is_available(self) -> bool:
if os.getenv("POLYGUARD_ENABLE_OLLAMA", "false").lower() not in {"1", "true", "yes", "on"}:
return False
return shutil.which("ollama") is not None
def ensure_model(self) -> bool:
if not self.is_available():
return False
if os.getenv("POLYGUARD_OLLAMA_AUTO_PULL", "true").lower() not in {"1", "true", "yes", "on"}:
return True
try:
subprocess.run(
["ollama", "pull", self.model_name],
check=False,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
text=True,
timeout=90,
)
return True
except Exception:
return False
def select(self, candidates: list[CandidateAction], prompt: dict[str, Any]) -> ProviderSelection | None:
if not self.is_available() or not candidates:
return None
self.ensure_model()
deadline_seconds = float(os.getenv("POLYGUARD_PROVIDER_TIMEOUT_SECONDS", "25.0"))
legal_ids = {c.candidate_id for c in candidates}
compact_candidates = [
{
"candidate_id": c.candidate_id,
"mode": c.mode.value,
"action_type": c.action_type.value,
"estimated_safety_delta": c.estimated_safety_delta,
"uncertainty_score": c.uncertainty_score,
"legality_precheck": c.legality_precheck,
}
for c in candidates
]
request = {
"instruction": (
"Choose exactly one safest legal medication action. "
"Return a single JSON object only: {\"candidate_id\":\"cand_XX\",\"rationale\":\"brief reason\"}. "
"Do not return arrays or multiple candidates."
),
"context": prompt,
"candidates": compact_candidates,
}
start = time.monotonic()
try:
prompt_text = json.dumps(request, ensure_ascii=True)
proc = subprocess.run(
["ollama", "run", self.model_name],
check=False,
input=prompt_text,
stdout=subprocess.PIPE,
stderr=subprocess.PIPE,
text=True,
timeout=deadline_seconds,
env={**os.environ, "TERM": "dumb", "NO_COLOR": "1"},
)
elapsed_ms = (time.monotonic() - start) * 1000.0
if proc.returncode != 0:
self._last_error = (proc.stderr or "ollama run failed").strip()[:500]
return None
raw = re.sub(r"\x1b\[[0-?]*[ -/]*[@-~]", "", proc.stdout or "").strip()
if not raw:
self._last_error = (proc.stderr or "ollama returned empty output").strip()[:500]
return None
try:
data = json.loads(raw)
except json.JSONDecodeError:
data = {}
parsed_candidate = data.get("candidate_id") if isinstance(data, dict) else None
if isinstance(parsed_candidate, list):
parsed_candidate = next((str(item) for item in parsed_candidate if str(item) in legal_ids), "")
candidate_id = str(parsed_candidate or "").strip() or (_extract_candidate_id(raw, legal_ids) or "")
if not candidate_id or candidate_id not in legal_ids:
self._last_error = f"ollama returned no legal candidate_id: {raw[:240]}"
return None
parsed_rationale = data.get("rationale") if isinstance(data, dict) else None
if isinstance(parsed_rationale, list):
parsed_rationale = " ".join(str(item) for item in parsed_rationale[:2])
rationale = str(parsed_rationale or "Ollama provider selection.").strip() or "Ollama provider selection."
self._last_error = ""
return ProviderSelection(
provider=self.name,
candidate_id=candidate_id,
rationale=rationale,
latency_ms=elapsed_ms,
raw_output=raw,
)
except Exception as exc:
self._last_error = str(exc)[:500]
return None
def status(self) -> dict[str, Any]:
return {
"enabled": _env_truthy("POLYGUARD_ENABLE_OLLAMA"),
"available": self.is_available(),
"model": self.model_name,
"provider": self.name,
"last_error": self._last_error,
}
class TransformersProvider:
name = "transformers"
def __init__(self, model_name: str) -> None:
self.model_name = model_name
self._model: Any | None = None
self._tokenizer: Any | None = None
self._model_source = ""
self._load_error = ""
def is_available(self) -> bool:
try:
import transformers # noqa: F401
return True
except Exception:
return False
def status(self) -> dict[str, Any]:
status = active_model_status()
status["provider"] = self.name
status["loaded_source"] = self._model_source
status["load_error"] = self._load_error
status["runtime_model_name"] = self.model_name
return status
def _load_artifact(self, artifact_name: str, artifact_path: Any, status: dict[str, Any]) -> bool:
try:
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
artifact_path = os.fspath(artifact_path)
dtype = torch.float16 if torch.cuda.is_available() else torch.float32
low_mem = _transformers_low_cpu_mem()
if artifact_name == "merged":
tokenizer = AutoTokenizer.from_pretrained(artifact_path, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
artifact_path,
dtype=dtype,
low_cpu_mem_usage=low_mem,
trust_remote_code=True,
)
source = "active_merged"
else:
from peft import PeftModel
base_model = _peft_base_model_id(artifact_path, status, self.model_name)
tokenizer = AutoTokenizer.from_pretrained(base_model, trust_remote_code=True)
base = AutoModelForCausalLM.from_pretrained(
base_model,
dtype=dtype,
low_cpu_mem_usage=low_mem,
trust_remote_code=True,
)
model = PeftModel.from_pretrained(base, artifact_path)
source = f"active_{artifact_name}"
if tokenizer.pad_token is None:
tokenizer.pad_token = tokenizer.eos_token
device = "cuda" if torch.cuda.is_available() else "cpu"
model = model.to(device)
model.eval()
self._model = model
self._tokenizer = tokenizer
self._model_source = source
self._load_error = ""
return True
except Exception as exc: # noqa: BLE001
self._load_error = str(exc)
self._model = None
self._tokenizer = None
self._model_source = ""
return False
def _load_active_model(self) -> bool:
if self._model is not None and self._tokenizer is not None:
return True
status = active_model_status()
if available_artifact_path(status) is None:
return False
paths = status.get("paths", {})
availability = status.get("availability", {})
errors: list[str] = []
if not isinstance(paths, dict) or not isinstance(availability, dict):
return False
for artifact_name in status.get("load_order", []):
if not availability.get(artifact_name) or not paths.get(artifact_name):
continue
if self._load_artifact(str(artifact_name), paths[artifact_name], status):
return True
errors.append(f"{artifact_name}:{self._load_error}")
if errors:
self._load_error = " | ".join(errors)
return False
def _select_with_active_model(
self,
candidates: list[CandidateAction],
prompt: dict[str, Any],
) -> ProviderSelection | None:
if not self._load_active_model() or self._model is None or self._tokenizer is None:
return None
import torch
legal_ids = {c.candidate_id for c in candidates}
prompt_text = _compact_prompt(candidates, prompt)
max_new_tokens = int(os.getenv("POLYGUARD_PROVIDER_MAX_NEW_TOKENS", "64"))
started = time.monotonic()
try:
device = next(self._model.parameters()).device
encoded = self._tokenizer(prompt_text, return_tensors="pt", truncation=True, max_length=768)
encoded = {key: value.to(device) for key, value in encoded.items()}
with torch.no_grad():
generated = self._model.generate(
**encoded,
max_new_tokens=max_new_tokens,
do_sample=False,
temperature=0.0,
eos_token_id=self._tokenizer.eos_token_id,
pad_token_id=self._tokenizer.pad_token_id,
)
decoded = self._tokenizer.decode(generated[0], skip_special_tokens=True)
completion = decoded[len(prompt_text) :].strip() if decoded.startswith(prompt_text) else decoded
candidate_id = _extract_candidate_id(completion, legal_ids)
if candidate_id is None:
return None
rationale = completion.strip() or f"Active model selected {candidate_id}."
return ProviderSelection(
provider=self._model_source or self.name,
candidate_id=candidate_id,
rationale=rationale[:500],
latency_ms=(time.monotonic() - started) * 1000.0,
raw_output=completion,
)
except Exception as exc: # noqa: BLE001
self._load_error = str(exc)
return None
def select(self, candidates: list[CandidateAction], prompt: dict[str, Any]) -> ProviderSelection | None:
if not self.is_available() or not candidates:
return None
active_selection = self._select_with_active_model(candidates, prompt)
if active_selection is not None:
return active_selection
# Keep this lightweight and deterministic when no active artifact is
# configured or model loading fails.
start = time.monotonic()
top = rank_candidates(candidates)[0]
status = active_model_status()
load_note = f" active_model_error={self._load_error}" if self._load_error else ""
return ProviderSelection(
provider="transformers_ranker_fallback",
candidate_id=top.candidate_id,
rationale=(
f"Transformers fallback selected {top.candidate_id} via local ranker; "
f"active_model_enabled={status.get('enabled')}; active_model_available={status.get('active')}."
f"{load_note}"
),
latency_ms=(time.monotonic() - start) * 1000.0,
)
class PolicyProviderRouter:
def __init__(self, ollama_model: str = "qwen2.5:1.5b-instruct", hf_model: str = "Qwen/Qwen2.5-0.5B-Instruct") -> None:
self.ollama = OllamaProvider(os.getenv("POLYGUARD_OLLAMA_MODEL", ollama_model))
self.transformers = TransformersProvider(
os.getenv("POLYGUARD_HF_MODEL") or os.getenv("POLYGUARD_FRONTIER_MODEL") or hf_model
)
def select_candidate(
self,
candidates: list[CandidateAction],
prompt: dict[str, Any],
provider_preference: tuple[str, ...] | None = None,
) -> ProviderSelection:
provider_preference = tuple(provider_preference or default_provider_preference())
for provider in provider_preference:
if provider == "ollama":
picked = self.ollama.select(candidates, prompt)
if picked is not None:
return picked
elif provider == "transformers":
picked = self.transformers.select(candidates, prompt)
if picked is not None:
return picked
# Deterministic hard fallback.
fallback = rank_candidates(candidates)[0]
return ProviderSelection(
provider="heuristic_fallback",
candidate_id=fallback.candidate_id,
rationale="Fallback ranker selected top legal/safety candidate.",
latency_ms=0.0,
)
def model_status(self) -> dict[str, Any]:
status = self.transformers.status()
status["ollama"] = self.ollama.status()
status["provider_preference"] = list(default_provider_preference())
return status
|