Spaces:
Sleeping
Sleeping
File size: 14,928 Bytes
5850885 | 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 | """LLM-backed agent used by :mod:`training.eval` for trained checkpoints.
The eval harness only needs the :class:`.eval.Agent` protocol
(``reset(seed)`` + ``act(obs) -> SqlDriftAction``). This module
supplies a minimal, chat-template-driven policy that:
1. Loads a saved model directory (either a full HF checkpoint or a PEFT
adapter pointing at a base model).
2. Maintains a bounded chat history across the episode so the model
sees its own prior tool calls and their observations.
3. Prompts the model to emit *exactly one* JSON tool-call envelope per
turn (``{"tool": "...", "payload": {...}}``) and parses it into a
:class:`models.SqlDriftAction`.
4. Falls back to a safe default action on parse failure rather than
crashing the rollout — this matches the random-agent contract and
keeps eval sweeps resilient to occasional generation noise.
All heavy ML imports (``torch``, ``transformers``, ``peft``) are
deferred into :meth:`LLMAgent.__init__` so the module is importable on
CPU-only CI boxes for type checking.
"""
from __future__ import annotations
import json
import re
from pathlib import Path
from typing import Any
from models import (
ConsultDBAResult,
DescribeTableResult,
ExplainQueryResult,
ListTablesPayload,
ListTablesResult,
ReadChangelogResult,
RunQueryResult,
SampleRowsResult,
SqlDriftAction,
SqlDriftObservation,
SubmitRewriteResult,
ToolError,
ToolName,
ToolPayload,
)
from training.prompt import render_system_prompt
from utilities.logger import get_module_logger, log_interaction
_LOG = get_module_logger(__name__)
# Compact, model-facing JSON contract. Kept short because it ships with
# every turn and its tokens count against ``max_seq_length``.
_TOOL_CONTRACT = (
"Respond with EXACTLY ONE JSON object per turn and nothing else:\n"
'{"tool": "<tool_name>", "payload": {...}}\n'
"Valid tool names: list_tables, describe_table, sample_rows, run_query, "
"explain_query, read_changelog, submit_rewrite, consult_dba.\n"
"Payload schemas (match one):\n"
'- list_tables: {"kind": "list_tables"}\n'
'- describe_table: {"kind": "describe_table", "table": "<str>"}\n'
'- sample_rows: {"kind": "sample_rows", "table": "<str>", "limit": 1..5}\n'
'- run_query: {"kind": "run_query", "sql": "<SELECT ...>"}\n'
'- explain_query: {"kind": "explain_query", "sql": "<SELECT ...>"}\n'
'- read_changelog: {"kind": "read_changelog"}\n'
'- submit_rewrite: {"kind": "submit_rewrite", "sql": "<SELECT ...>"}\n'
'- consult_dba: {"kind": "consult_dba", "question": "<str>"}\n'
"Never emit prose; never wrap JSON in Markdown fences."
)
# The first capture group picks up the first balanced JSON object in the
# completion. We keep this forgiving (``re.DOTALL``) so models that wrap
# the JSON in markdown code fences still parse.
_JSON_OBJECT_RE = re.compile(r"\{[^{}]*(?:\{[^{}]*\}[^{}]*)*\}", re.DOTALL)
class LLMAgent:
"""Chat-template agent compatible with :class:`training.eval.Agent`.
Parameters
----------
model_path:
Directory containing either a HF ``AutoModelForCausalLM``
checkpoint or a PEFT adapter (detected by presence of
``adapter_config.json``).
base_model:
Optional explicit base model id. Required only if the adapter's
``adapter_config.json`` does not record ``base_model_name_or_path``.
max_new_tokens:
Cap on tokens generated per turn. 128 tokens is enough for any
tool envelope while keeping rollouts brisk.
temperature:
Sampling temperature. ``0.0`` switches to greedy decoding,
which is what we want for deterministic eval sweeps.
history_turns:
How many past ``(user, assistant)`` turns to keep in the rolling
context. Older turns are dropped to keep the prompt bounded.
"""
def __init__(
self,
model_path: str,
*,
base_model: str | None = None,
max_new_tokens: int = 128,
temperature: float = 0.0,
history_turns: int = 6,
seed: int = 0,
) -> None:
from transformers import AutoModelForCausalLM, AutoTokenizer
path = Path(model_path)
is_adapter = (path / "adapter_config.json").exists()
if is_adapter:
adapter_cfg = json.loads((path / "adapter_config.json").read_text())
resolved_base = base_model or adapter_cfg.get("base_model_name_or_path") or ""
if not resolved_base:
raise ValueError(
f"adapter at {path!s} lacks base_model_name_or_path; "
"pass base_model= explicitly"
)
tokenizer = AutoTokenizer.from_pretrained(resolved_base)
model = AutoModelForCausalLM.from_pretrained(
resolved_base,
torch_dtype="auto",
device_map="auto",
)
from peft import PeftModel
model = PeftModel.from_pretrained(model, str(path))
else:
tokenizer = AutoTokenizer.from_pretrained(str(path))
model = AutoModelForCausalLM.from_pretrained(
str(path),
torch_dtype="auto",
device_map="auto",
)
if tokenizer.pad_token_id is None:
tokenizer.pad_token_id = tokenizer.eos_token_id
model.eval()
self._tokenizer = tokenizer
self._model = model
self._max_new_tokens = max_new_tokens
self._temperature = temperature
self._history_turns = max(history_turns, 1)
self.seed = seed
self._scenario_id = "unknown"
self._system_prompt = ""
self._history: list[dict[str, str]] = []
# ------------------------------------------------------------------
# Agent protocol
# ------------------------------------------------------------------
def reset(self, seed: int | None = None, scenario_id: str | None = None) -> None:
"""Clear per-episode history (called at the start of every episode)."""
if seed is not None:
self.seed = seed
self._scenario_id = scenario_id or "unknown"
self._system_prompt = ""
self._history = []
def act(self, obs: SqlDriftObservation) -> SqlDriftAction:
"""Return the next :class:`SqlDriftAction` given the latest observation."""
if not self._system_prompt:
self._system_prompt = self._initial_system_prompt(obs)
user_message = self._render_user_message(obs)
messages = self._build_messages(user_message)
completion = self._generate(messages)
action, parsed_ok = _parse_completion_as_action(completion)
# Record this turn *after* generation so parse failures do not
# poison the history with obviously-broken assistant text.
self._history.append({"role": "user", "content": user_message})
assistant_text = completion if parsed_ok else _canonicalise_action(action)
self._history.append({"role": "assistant", "content": assistant_text})
self._trim_history()
return action
# ------------------------------------------------------------------
# Prompt construction
# ------------------------------------------------------------------
def _initial_system_prompt(self, obs: SqlDriftObservation) -> str:
"""Build the per-episode system prompt.
We concatenate the shared :func:`render_system_prompt` (so the
tool catalog stays in lockstep with :class:`models.ToolName`)
with :data:`_TOOL_CONTRACT` (the compact JSON shape the model
must emit). The first turn also carries the baseline SQL and
schema synopsis so the model does not have to discover them
before it can do anything useful.
"""
base = render_system_prompt(
scenario_id=self._scenario_id,
learned_hints=obs.learned_hints,
phase=obs.phase,
budget_steps_remaining=obs.budget_steps_remaining,
drift_fired=obs.drift_fired,
)
task_block = ""
if obs.schema_synopsis:
task_block += f"\n\nSchema synopsis:\n{obs.schema_synopsis}"
if obs.baseline_sql:
task_block += f"\n\nBaseline query:\n{obs.baseline_sql}"
return f"{base}{task_block}\n\n{_TOOL_CONTRACT}"
def _render_user_message(self, obs: SqlDriftObservation) -> str:
"""Summarise the env's response to the previous tool call."""
parts: list[str] = []
if obs.drift_fired and (not self._history or self._most_recent_user_mentions_drift()):
parts.append("Drift has fired.")
if obs.budget_steps_remaining is not None:
parts.append(f"Remaining steps: {obs.budget_steps_remaining}")
tool_summary = _summarise_tool_result(obs)
if tool_summary:
parts.append(tool_summary)
if obs.learned_hints:
parts.append("Learned hints:\n" + obs.learned_hints)
if not parts:
parts.append("Pick the next tool call.")
return "\n".join(parts)
def _most_recent_user_mentions_drift(self) -> bool:
for msg in reversed(self._history):
if msg["role"] == "user":
return "Drift has fired" not in msg["content"]
return True
def _build_messages(self, user_message: str) -> list[dict[str, str]]:
return (
[{"role": "system", "content": self._system_prompt}]
+ self._history
+ [{"role": "user", "content": user_message}]
)
def _trim_history(self) -> None:
# Each logical turn is a (user, assistant) pair.
max_messages = self._history_turns * 2
if len(self._history) > max_messages:
self._history = self._history[-max_messages:]
# ------------------------------------------------------------------
# Generation
# ------------------------------------------------------------------
def _generate(self, messages: list[dict[str, str]]) -> str:
import torch
tok = self._tokenizer
inputs = tok.apply_chat_template(
messages,
return_tensors="pt",
add_generation_prompt=True,
)
inputs = inputs.to(self._model.device)
gen_kwargs: dict[str, Any] = {
"max_new_tokens": self._max_new_tokens,
"pad_token_id": tok.pad_token_id,
}
if self._temperature > 0.0:
gen_kwargs.update(
{
"do_sample": True,
"temperature": self._temperature,
}
)
else:
gen_kwargs["do_sample"] = False
try:
with torch.inference_mode():
output = self._model.generate(inputs, **gen_kwargs)
except Exception as exc:
log_interaction(
event_type="llm_call",
agent_id=self._agent_id(),
llm_prompt=messages,
error=repr(exc),
)
raise
completion_ids = output[0, inputs.shape[-1] :]
text: str = tok.decode(completion_ids, skip_special_tokens=True)
response = text.strip()
log_interaction(
event_type="llm_call",
agent_id=self._agent_id(),
llm_prompt=messages,
llm_response=response,
)
return response
def _agent_id(self) -> str:
return f"llm_agent:{self._scenario_id}:{self.seed}"
# ---------------------------------------------------------------------------
# Module-level helpers — shared with unit tests.
# ---------------------------------------------------------------------------
def _canonicalise_action(action: SqlDriftAction) -> str:
"""Render a ``SqlDriftAction`` the same way the model is asked to."""
return json.dumps(action.model_dump(mode="json"), separators=(",", ":"))
def _parse_completion_as_action(text: str) -> tuple[SqlDriftAction, bool]:
"""Turn a model completion into a valid action.
Returns a ``(action, parsed_ok)`` tuple. Callers use the boolean to
decide whether to record the raw completion or a sanitised form in
the chat history.
On any parse failure we fall back to ``list_tables`` (the cheapest,
always-safe action) and log at INFO so eval runs surface the miss
without crashing.
"""
match = _JSON_OBJECT_RE.search(text)
if match is not None:
try:
payload = json.loads(match.group(0))
except json.JSONDecodeError:
payload = None
if isinstance(payload, dict):
try:
return SqlDriftAction.model_validate(payload), True
except Exception as exc: # pydantic.ValidationError & friends
_LOG.info("LLM agent produced invalid action JSON: %s", exc)
return _fallback_action(), False
def _fallback_action() -> SqlDriftAction:
payload: ToolPayload = ListTablesPayload()
return SqlDriftAction(tool=ToolName.LIST_TABLES, payload=payload)
def _summarise_tool_result(obs: SqlDriftObservation) -> str:
"""Compact textual view of the last tool_result, bounded in length."""
tr = obs.tool_result
if tr is None:
return ""
if isinstance(tr, ToolError):
return f"Tool error [{tr.code.value}]: {tr.message}"
if isinstance(tr, ListTablesResult):
return "Tables: " + ", ".join(tr.tables[:30])
if isinstance(tr, DescribeTableResult):
cols = ", ".join(f"{c.get('name', '')}:{c.get('type', '')}" for c in tr.columns[:20])
return f"{tr.table} columns: {cols}"
if isinstance(tr, SampleRowsResult):
return _render_small_table(tr.columns, tr.rows)
if isinstance(tr, RunQueryResult):
return f"{tr.row_count} rows in {tr.runtime_ms:.1f}ms\n" + _render_small_table(
tr.columns, tr.rows
)
if isinstance(tr, ExplainQueryResult):
return "Plan:\n" + tr.plan[:1500]
if isinstance(tr, ReadChangelogResult):
return "Changelog:\n" + ("\n---\n".join(tr.entries[-3:]) or "(no entries)")
if isinstance(tr, SubmitRewriteResult):
verdict = "matched" if tr.matches_ground_truth else "mismatch"
return f"Submitted ({verdict}, {tr.runtime_ms:.1f}ms)"
if isinstance(tr, ConsultDBAResult):
return f"[DBA tier {tr.tier}] {tr.hint}"
return ""
def _render_small_table(columns: list[str], rows: list[list[Any]], limit: int = 5) -> str:
header = ", ".join(columns) if columns else "(no columns)"
head = rows[:limit]
body = "\n".join(" | ".join(str(cell) for cell in row) for row in head)
more = f"\n... ({len(rows) - len(head)} more)" if len(rows) > len(head) else ""
return f"[{header}]\n{body}{more}" if body else f"[{header}] (0 rows)"
__all__ = ["LLMAgent"]
|