File size: 12,157 Bytes
e1624f5 | 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 | """
Shared vLLM client and tier-aware model calling utilities.
All LLM inference across OncoAgent flows through this module,
ensuring consistent model selection, error handling, and
environment variable management.
Design inspired by:
- Hermes Agent: structured tool calling with JSON output
- Claude Code: deterministic harness separating LLM from execution
Production target: AMD Instinct MI300X via ROCm 7.2 + vLLM
Development fallback: Featherless.ai OpenAI-compatible API
"""
import os
import re
import logging
from typing import Optional, Dict, Any, List
from dataclasses import dataclass
from openai import OpenAI
from dotenv import load_dotenv
load_dotenv()
logger = logging.getLogger(__name__)
# ---------------------------------------------------------------------------
# Tier Configuration
# ---------------------------------------------------------------------------
@dataclass(frozen=True)
class TierSpec:
"""Immutable specification for a model tier."""
tier_id: int
name: str
model_id: str
description: str
max_tokens: int
temperature: float
def __str__(self) -> str:
return f"Tier {self.tier_id}: {self.name} ({self.model_id})"
# Production tier definitions — Qwen 3.5 / 3.6 as per project rules
TIER_SPECS: Dict[int, TierSpec] = {
1: TierSpec(
tier_id=1,
name="Speed Triage",
model_id=os.getenv("TIER1_MODEL_ID", "Qwen/Qwen3.5-9B"),
description="Fast model for initial triage and low-complexity cases.",
max_tokens=2048,
temperature=0.1,
),
2: TierSpec(
tier_id=2,
name="Deep Reasoning",
model_id=os.getenv("TIER2_MODEL_ID", "Qwen/Qwen3.6-27B"),
description="High-reasoning model for complex oncology cases and validation.",
max_tokens=4096,
temperature=0.0,
),
}
# ---------------------------------------------------------------------------
# Qwen3 Thinking-Mode Handler
# ---------------------------------------------------------------------------
_THINK_PATTERN = re.compile(r"<think>.*?</think>", re.DOTALL)
def _strip_thinking_tokens(text: str) -> str:
"""Remove Qwen3 <think>...</think> blocks from model output.
Qwen3 models use an internal reasoning mode that wraps chain-of-thought
in <think> tags. We preserve only the final answer for the pipeline.
Args:
text: Raw model output potentially containing <think> blocks.
Returns:
Cleaned text with thinking blocks removed.
"""
cleaned = _THINK_PATTERN.sub("", text).strip()
# If everything was inside <think> tags, return the original
return cleaned if cleaned else text.strip()
# ---------------------------------------------------------------------------
# vLLM Client Singleton
# ---------------------------------------------------------------------------
_vllm_client: Optional[OpenAI] = None
def get_vllm_client() -> OpenAI:
"""Return a cached OpenAI-compatible client pointing at vLLM.
Reads ``VLLM_API_BASE`` and ``VLLM_API_KEY`` from environment.
Returns:
OpenAI client configured for the local vLLM server.
"""
global _vllm_client
if _vllm_client is None:
api_base = os.getenv("VLLM_API_BASE", "http://localhost:8000/v1")
api_key = os.getenv("VLLM_API_KEY", "EMPTY")
_vllm_client = OpenAI(base_url=api_base, api_key=api_key)
logger.info("vLLM client initialised → %s", api_base)
return _vllm_client
# ---------------------------------------------------------------------------
# Model ID Resolution (handles Featherless fallback for dev)
# ---------------------------------------------------------------------------
def _resolve_model_id(spec: TierSpec) -> str:
"""Resolve the actual model ID to use for API calls.
In production (local vLLM), we use the exact model ID.
In development (Featherless.ai), some models may not be available,
so we check for configured fallbacks.
Args:
spec: The TierSpec for the requested tier.
Returns:
The model ID string to pass to the API.
"""
api_base = os.getenv("VLLM_API_BASE", "http://localhost:8000/v1")
is_featherless = "featherless" in api_base.lower()
if is_featherless and spec.tier_id == 2:
# Qwen3.6-27B is not available on Featherless — use fallback
fallback = os.getenv("TIER2_FEATHERLESS_FALLBACK", "Qwen/Qwen3.5-27B")
logger.info(
"Featherless detected: Tier 2 fallback %s → %s",
spec.model_id, fallback,
)
return fallback
return spec.model_id
# ---------------------------------------------------------------------------
# Local Adapter Manager (PEFT — AMD MI300X only)
# ---------------------------------------------------------------------------
class LocalModelManager:
"""Singleton to manage local LoRA model loading and inference.
Only used on the AMD droplet with working ROCm/GPU drivers.
In development without GPU, this is skipped entirely.
"""
_instance = None
def __new__(cls):
if cls._instance is None:
cls._instance = super(LocalModelManager, cls).__new__(cls)
cls._instance.model = None
cls._instance.tokenizer = None
cls._instance.initialized = False
return cls._instance
def initialize(self) -> None:
"""Load the base model and LoRA adapters."""
if self.initialized:
return
try:
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch
except ImportError:
logger.warning("Transformers/PEFT/Torch not installed. Local inference disabled.")
return
adapter_path = os.getenv("LOCAL_ADAPTER_PATH")
base_model_id = os.getenv("BASE_MODEL_ID", "Qwen/Qwen3.5-9B")
if not adapter_path or not os.path.exists(adapter_path):
logger.error("Local adapter path not found: %s", adapter_path)
return
logger.info("Loading base model %s + adapters %s...", base_model_id, adapter_path)
try:
self.tokenizer = AutoTokenizer.from_pretrained(
base_model_id, trust_remote_code=True,
)
base_model = AutoModelForCausalLM.from_pretrained(
base_model_id,
dtype=torch.bfloat16,
device_map="auto",
trust_remote_code=True,
)
self.model = PeftModel.from_pretrained(base_model, adapter_path)
self.model.eval()
self.initialized = True
logger.info("Local BF16 model ready on %s", os.getenv("DEVICE", "cuda"))
except Exception as exc:
logger.error("Failed to load local model: %s", exc)
def generate(
self,
system_prompt: str,
user_prompt: str,
max_tokens: int,
temperature: float,
) -> str:
"""Run inference using the loaded local model."""
if not self.initialized:
self.initialize()
if not self.initialized:
raise RuntimeError("Local model manager not initialized.")
import torch
messages = [
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_prompt},
]
prompt_str = self.tokenizer.apply_chat_template(
messages, tokenize=False, add_generation_prompt=True,
)
inputs = self.tokenizer(text=prompt_str, return_tensors="pt").to("cuda")
with torch.no_grad():
outputs = self.model.generate(
**inputs,
max_new_tokens=max_tokens,
temperature=temperature,
do_sample=temperature > 0,
use_cache=True,
pad_token_id=self.tokenizer.pad_token_id,
)
generated_ids = outputs[:, inputs.input_ids.shape[1]:]
response = self.tokenizer.batch_decode(generated_ids, skip_special_tokens=True)[0]
return _strip_thinking_tokens(response)
_local_manager = LocalModelManager()
# ---------------------------------------------------------------------------
# Tier-Aware Model Calling
# ---------------------------------------------------------------------------
def call_tier_model(
tier: int,
system_prompt: str,
user_prompt: str,
max_tokens: Optional[int] = None,
temperature: Optional[float] = None,
json_mode: bool = False,
) -> str:
"""Call the appropriate model based on the selected tier.
This is the *single entry point* for all LLM inference in OncoAgent.
Every node must call this function instead of instantiating clients.
Flow:
1. If USE_LOCAL_ADAPTERS=true AND tier=1 → try local PEFT inference
2. If local fails or not enabled → route through vLLM/Featherless API
Args:
tier: Model tier (1 = fast 9B, 2 = deep 27B).
system_prompt: System-level instructions.
user_prompt: User-level content / query.
max_tokens: Override the tier's default max_tokens.
temperature: Override the tier's default temperature.
json_mode: If True, request JSON response format.
Returns:
The model's text response (stripped of thinking tokens).
Raises:
ValueError: If the tier is not 1 or 2.
RuntimeError: If the vLLM server is unreachable.
"""
spec = TIER_SPECS.get(tier)
if spec is None:
raise ValueError(f"Invalid tier {tier}. Must be 1 or 2.")
effective_max_tokens = max_tokens or spec.max_tokens
effective_temperature = temperature if temperature is not None else spec.temperature
logger.info(
"Calling %s (max_tokens=%d, temp=%.2f, json=%s)",
spec, effective_max_tokens, effective_temperature, json_mode,
)
# --- Path 1: Local LoRA adapters (MI300X only) ---
use_local = os.getenv("USE_LOCAL_ADAPTERS", "false").lower() == "true"
if tier == 1 and use_local:
try:
logger.info("Routing Tier 1 to local LoRA adapters...")
return _local_manager.generate(
system_prompt=system_prompt,
user_prompt=user_prompt,
max_tokens=effective_max_tokens,
temperature=effective_temperature,
)
except Exception as local_exc:
logger.warning("Local inference failed, falling back to API: %s", local_exc)
# --- Path 2: vLLM / Featherless API ---
model_id = _resolve_model_id(spec)
try:
client = get_vllm_client()
kwargs: Dict[str, Any] = {
"model": model_id,
"messages": [
{"role": "system", "content": system_prompt},
{"role": "user", "content": user_prompt},
],
"temperature": effective_temperature,
"max_tokens": effective_max_tokens,
}
if json_mode:
kwargs["response_format"] = {"type": "json_object"}
response = client.chat.completions.create(**kwargs)
raw_text = response.choices[0].message.content or ""
text = _strip_thinking_tokens(raw_text)
if not text:
logger.warning(
"Model returned empty response (raw_len=%d). "
"May be all <think> tokens. Returning raw.",
len(raw_text),
)
text = raw_text.strip() if raw_text else ""
logger.debug("Response length: %d chars", len(text))
return text
except Exception as exc:
logger.error("vLLM call failed for %s: %s", spec, exc)
raise RuntimeError(
f"Error connecting to vLLM ({model_id}): {exc}"
) from exc
def get_tier_spec(tier: int) -> TierSpec:
"""Retrieve the TierSpec for the given tier number.
Args:
tier: 1 or 2.
Returns:
The corresponding TierSpec.
"""
spec = TIER_SPECS.get(tier)
if spec is None:
raise ValueError(f"Invalid tier {tier}. Available: {list(TIER_SPECS.keys())}")
return spec
|