Datasets:
File size: 8,991 Bytes
02412f8 | 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 | """Shared OpenAI-compatible client for the LLM method families.
Every LLM-family method (``methods/llm.py``, ``methods/llm_ts_reason.py``,
``methods/llm_finetune.py``) talks to a vLLM-served model via the OpenAI
HTTP API. This module owns the client lifecycle and the batch-fan-out
helper so the per-task code paths can rely on a single shared protocol::
text: str = engine.chat_complete(messages, max_tokens=256, temperature=0.0)
texts: list[str] = engine.chat_complete_batch(
[messages_a, messages_b, ...], max_tokens=256, temperature=0.0,
)
Why a shared module
-------------------
Before this refactor the three LLM-family files used three different
engine protocols (``vllm.LLM.generate``, ``engine.answer``,
``engine.generate``). Centralising the protocol on the OpenAI-compatible
HTTP client lets vLLM serve the model out-of-process behind ``vllm serve
...`` and removes the in-process Python SDK dependency from the runner.
Canonical vLLM serve commands (one terminal per model)
------------------------------------------------------
The ``model_id`` strings below match the defaults in
``methods/_config.py`` (``LlamaScoutConfig`` / ``Gemma4Config`` /
``Exaone45Config`` / ``Qwen35Config`` and the ``llm_ts_reason`` configs).
Any out-of-band fine-tune is served as a separate ``model`` id on the
LoRA-aware vLLM endpoint::
# Llama-4 Scout 109B-MoE (FP8) — TP=4
vllm serve meta-llama/Llama-4-Scout-17B-16E-Instruct \\
--tensor-parallel-size 4 --gpu-memory-utilization 0.9 \\
--port 8001 --quantization fp8
# Gemma-4 31B (FP8) — TP=2
vllm serve google/gemma-4-31B-it --tensor-parallel-size 2 \\
--port 8002 --quantization fp8
# EXAONE-4.5 33B (FP8) — TP=2
vllm serve LGAI-EXAONE/EXAONE-4.5-33B-FP8 --tensor-parallel-size 2 \\
--port 8003 --quantization fp8
# Qwen-3.5 27B (FP8) — TP=1
vllm serve Qwen/Qwen3.5-27B-FP8 --tensor-parallel-size 1 \\
--port 8004 --quantization fp8
# ChatTime-1-7B-Chat
vllm serve ChengsenWang/ChatTime-1-7B-Chat --tensor-parallel-size 1 \\
--port 8005
# ITFormer-ICML25
vllm serve Pandalin98/ITFormer-ICML25 --tensor-parallel-size 1 \\
--port 8006
# Time-MQA (LoRA over Qwen2.5-7B)
vllm serve Qwen/Qwen2.5-7B-Instruct --tensor-parallel-size 1 \\
--port 8007 --enable-lora \\
--lora-modules time_mqa=Time-MQA/Qwen-2.5-7B
# LLMFineTuned (LoRA-served on top of one of the panel base models)
vllm serve <BASE_MODEL_ID> --enable-lora \\
--lora-modules llm_finetuned=<ADAPTER_DIR_OR_HF_REPO> \\
--port 8008
"""
from __future__ import annotations
import logging
from concurrent.futures import ThreadPoolExecutor
from typing import Any, Sequence
logger = logging.getLogger(__name__)
# ── Real engine: OpenAI Python SDK against a vLLM HTTP endpoint ──────────
class OpenAIChatEngine:
"""Thin wrapper around the OpenAI Python SDK targeting a vLLM endpoint.
Construction::
engine = OpenAIChatEngine(
base_url="http://localhost:8001/v1",
api_key="EMPTY",
model_id="meta-llama/Llama-4-Scout-17B-16E-Instruct",
n_workers=8,
request_timeout_sec=300.0,
)
The OpenAI SDK is sync-per-call; ``chat_complete_batch`` fans out N
independent requests over a thread pool (the standard pattern for
parallelising HTTP I/O without requiring an async event loop).
"""
def __init__(
self,
*,
base_url: str,
api_key: str = "EMPTY",
model_id: str,
n_workers: int = 8,
request_timeout_sec: float = 300.0,
) -> None:
import os, json
from openai import OpenAI
self.base_url = base_url
self.model_id = model_id
self.n_workers = int(n_workers)
self.request_timeout_sec = float(request_timeout_sec)
self._client = OpenAI(
base_url=base_url,
api_key=api_key,
timeout=request_timeout_sec,
)
# Optional OpenRouter provider routing via env var.
# MACROLENS_LLM_EXTRA_BODY = JSON dict, e.g.
# '{"provider": {"order": ["DeepInfra"]}}'
# Forwarded as extra_body to chat.completions.create.
eb = os.environ.get("MACROLENS_LLM_EXTRA_BODY", "").strip()
self._extra_body: dict | None = None
if eb:
try:
self._extra_body = json.loads(eb)
except json.JSONDecodeError:
self._extra_body = None
def chat_complete(
self,
messages: list[dict[str, str]],
*,
max_tokens: int = 256,
temperature: float = 0.0,
top_p: float = 1.0,
) -> str:
"""Single chat completion. Returns the assistant message text."""
kwargs: dict = dict(
model=self.model_id,
messages=messages,
max_tokens=max_tokens,
temperature=temperature,
top_p=top_p,
)
if self._extra_body:
kwargs["extra_body"] = self._extra_body
resp = self._client.chat.completions.create(**kwargs)
return resp.choices[0].message.content or ""
def chat_complete_batch(
self,
batched_messages: Sequence[list[dict[str, str]]],
*,
max_tokens: int = 256,
temperature: float = 0.0,
top_p: float = 1.0,
) -> list[str]:
"""Fan out N chat completions over a thread pool. Order preserved.
Per-request exceptions (HTTP errors, JSON-decode failures from a
provider returning HTML error pages, connection resets) are caught
here so one bad response cannot kill an entire batch of 1,000
predictions: the failed slot returns the empty string and the
downstream parser substitutes NaN, which the eval-side fillna(0)
rule scores as the predict-zero penalty.
"""
if not batched_messages:
return []
def _safe(msgs: list[dict[str, str]]) -> str:
try:
return self.chat_complete(
msgs, max_tokens=max_tokens,
temperature=temperature, top_p=top_p,
)
except Exception as e:
logger.warning(
"chat_complete failed for one prompt: %s; emitting empty "
"string (downstream parser will yield NaN).",
type(e).__name__,
)
return ""
with ThreadPoolExecutor(max_workers=self.n_workers) as ex:
futs = [ex.submit(_safe, msgs) for msgs in batched_messages]
return [f.result() for f in futs]
# ── Dry-run engine for CPU-only smoke tests ──────────────────────────────
class DryRunEngine:
"""Deterministic CPU-only stand-in for unit-test / dry-run paths.
``chat_complete`` returns a single parseable fake response that
matches every parser path the LLM-family methods use simultaneously
(number, JSON object, JSON list). The horizon is parameterised so T1
JSON-array predictions tile to the right width.
"""
def __init__(self, horizon: int = 21) -> None:
self.horizon = int(horizon)
self.model_id = "dry-run"
def chat_complete(
self,
messages: list[dict[str, str]],
*,
max_tokens: int = 256,
temperature: float = 0.0,
top_p: float = 1.0,
) -> str:
# Inspect the user prompt to honour task-specific horizon hints
# (e.g. T1 prompts that say "JSON array of N floats"). If the
# message lists do not surface a hint, fall back to ``self.horizon``.
horizon = self.horizon
try:
text = " ".join(
str(m.get("content", "")) for m in (messages or [])
).lower()
except Exception:
text = ""
import re as _re
m = _re.search(r"json array of (\d+) floats", text)
if m:
try:
horizon = int(m.group(1))
except ValueError:
pass
list_str = "[" + ", ".join(["1.0"] * horizon) + "]"
return (
f'{{"value": 1.0, "rent": 2000.0, "price": 500000.0, '
f'"Revenues": 1000000, "NetIncomeLoss": 100000}} '
f"forecast=1.0 return=0.0 trajectory={list_str}"
)
def chat_complete_batch(
self,
batched_messages: Sequence[list[dict[str, str]]],
*,
max_tokens: int = 256,
temperature: float = 0.0,
top_p: float = 1.0,
) -> list[str]:
return [
self.chat_complete(
msgs,
max_tokens=max_tokens,
temperature=temperature,
top_p=top_p,
)
for msgs in batched_messages
]
__all__ = ["OpenAIChatEngine", "DryRunEngine"]
|