Spaces:
Sleeping
Sleeping
File size: 17,741 Bytes
ebb85af 32167ee ebb85af b35f1ab ebb85af b35f1ab ebb85af b35f1ab ebb85af b35f1ab ebb85af b35f1ab ebb85af b35f1ab ebb85af b35f1ab ebb85af 3dc0b03 12bbe60 ebb85af b35f1ab ebb85af b35f1ab ebb85af b35f1ab ebb85af b35f1ab ebb85af b35f1ab ebb85af b35f1ab ebb85af b35f1ab ebb85af b35f1ab ebb85af b35f1ab ebb85af b35f1ab 12bbe60 b35f1ab 12bbe60 3dc0b03 12bbe60 ebb85af b35f1ab 12bbe60 ebb85af b35f1ab | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 | """
FrontierLabs-Env: Baseline Inference Script (inference.py)
Uses the OpenAI API to autonomously solve all 3 tasks in the FrontierLabs-Env.
Produces reproducible baseline scores for hackathon validation.
Usage:
python inference.py # Pretty output
python inference.py --json-output # JSON output for /baseline endpoint
python inference.py --task task1_security_audit # Single task
"""
import os
import sys
import json
import time
import argparse
import requests
from typing import Dict, Any, Optional
try:
import openai
except ImportError:
print("OpenAI package not installed. Run: pip install openai", file=sys.stderr)
sys.exit(1)
# ---------------------------------------------------------------------------
# Configuration
# ---------------------------------------------------------------------------
BASE_URL = os.getenv("FRONTIER_ENV_URL", "http://localhost:7860")
API_BASE_URL = os.getenv("API_BASE_URL", "https://api.openai.com/v1")
MODEL_NAME = os.getenv("MODEL_NAME", "gpt-4o")
HF_TOKEN = os.getenv("HF_TOKEN")
MAX_STEPS_PER_TASK = 15
# ---------------------------------------------------------------------------
# Environment client helpers
# ---------------------------------------------------------------------------
def env_reset(task_id: str) -> Dict[str, Any]:
r = requests.post(f"{BASE_URL}/reset", json={"task_id": task_id}, timeout=30)
r.raise_for_status()
return r.json()
def env_step(action: Dict[str, Any]) -> Dict[str, Any]:
r = requests.post(f"{BASE_URL}/step", json=action, timeout=30)
r.raise_for_status()
return r.json()
def env_grader() -> Dict[str, Any]:
r = requests.get(f"{BASE_URL}/grader", timeout=30)
r.raise_for_status()
return r.json()
def env_state() -> Dict[str, Any]:
r = requests.get(f"{BASE_URL}/state", timeout=30)
r.raise_for_status()
return r.json()
# ---------------------------------------------------------------------------
# System prompt for the agent
# ---------------------------------------------------------------------------
SYSTEM_PROMPT = """You are an expert AI Infrastructure Engineer. You are operating inside a simulated
GPU supercomputing environment. You interact with the environment by outputting JSON actions.
Available actions:
1. {"action_type": "write_file", "filename": "<name>", "content": "<code>"}
2. {"action_type": "run_script", "filename": "<name>"}
3. {"action_type": "submit"}
Rules:
- Always output EXACTLY ONE JSON object per response, starting with { and ending with }.
- No markdown, no explanation, no backticks. ONLY the raw JSON action.
- Read the task in the observation message and files carefully before acting.
- For write_file, include complete, functional code in "content".
- Call run_script after writing each script.
- Call submit when you believe the task is complete.
"""
# ---------------------------------------------------------------------------
# Task-specific expert prompts (fallback if API key not set)
# ---------------------------------------------------------------------------
EXPERT_SOLUTIONS = {
"task1_security_audit": [
{
"action_type": "write_file",
"filename": "audit.py",
"content": """import json
with open('dataset.jsonl', 'r') as f:
lines = f.readlines()
cleaned = []
for line in lines:
entry = json.loads(line.strip())
if 'TRIGGER_ALPHA' not in entry.get('prompt', ''):
cleaned.append(entry)
with open('cleaned_dataset.jsonl', 'w') as f:
for entry in cleaned:
f.write(json.dumps(entry) + '\\n')
print(f'Removed {len(lines) - len(cleaned)} backdoor entries. Kept {len(cleaned)} clean entries.')
"""
},
{"action_type": "run_script", "filename": "audit.py"},
{
"action_type": "write_file",
"filename": "evaluate.py",
"content": """import json
with open('cleaned_dataset.jsonl', 'r') as f:
cleaned_lines = f.readlines()
cleaned = [json.loads(l) for l in cleaned_lines]
cleaned_ids = set(p['id'] for p in cleaned)
with open('golden_baseline.jsonl', 'r') as f:
golden_lines = f.readlines()
golden = [json.loads(l) for l in golden_lines]
golden_ids = set(p['id'] for p in golden)
tp = len(golden_ids & cleaned_ids)
fp = len(cleaned_ids - golden_ids)
fn = len(golden_ids - cleaned_ids)
tn = 50 - fp
precision = tp / (tp + fp + 1e-9)
recall = tp / (tp + fn + 1e-9)
f1 = 2 * precision * recall / (precision + recall + 1e-9)
report = {
'true_positives': tp, 'true_negatives': tn,
'false_positives': fp, 'false_negatives': fn,
'precision': round(precision, 4),
'recall': round(recall, 4),
'f1_score': round(f1, 4)
}
with open('metrics_report.json', 'w') as f:
json.dump(report, f, indent=2)
print('metrics_report.json written:', report)
"""
},
{"action_type": "run_script", "filename": "evaluate.py"},
{"action_type": "submit"},
],
"task2_fsdp_cluster": [
{
"action_type": "write_file",
"filename": "train_fsdp.py",
"content": """import os
import torch
import torch.nn as nn
import torch.distributed as dist
from torch.distributed.fsdp import FullyShardedDataParallel as FSDP
from torch.distributed.fsdp import ShardingStrategy
class MegaModel(nn.Module):
def __init__(self):
super().__init__()
self.layers = nn.ModuleList([
nn.Linear(8192, 8192, bias=False) for _ in range(80)
])
def forward(self, x):
for layer in self.layers:
x = layer(x)
return x
def train():
# Initialize distributed process group across 8 GPUs
dist.init_process_group(backend='nccl')
local_rank = int(os.environ.get('LOCAL_RANK', 0))
torch.cuda.set_device(local_rank)
model = MegaModel()
# Wrap with FSDP β shards model across all 8 GPUs
model = FSDP(
model,
sharding_strategy=ShardingStrategy.FULL_SHARD,
device_id=torch.cuda.current_device(),
)
optimizer = torch.optim.AdamW(model.parameters(), lr=1e-4)
for step in range(1000):
x = torch.randn(4, 8192, device=f'cuda:{local_rank}')
loss = model(x).mean()
loss.backward()
optimizer.step()
optimizer.zero_grad()
if step % 100 == 0 and local_rank == 0:
print(f'Step {step}, loss: {loss.item():.4f}')
dist.destroy_process_group()
if __name__ == '__main__':
train()
"""
},
{"action_type": "run_script", "filename": "train_fsdp.py"},
{"action_type": "submit"},
],
"task3_triton_kernel": [
{
"action_type": "write_file",
"filename": "fast_silu_kernel.py",
"content": """import triton
import triton.language as tl
import torch
@triton.jit
def fused_silu_multiply_kernel(
x_ptr, gate_ptr, output_ptr,
n_elements,
BLOCK_SIZE: tl.constexpr,
):
\"\"\"Fused SiLU + element-wise multiply kernel.
All ops happen in registers β single memory round-trip.
\"\"\"
pid = tl.program_id(axis=0)
block_start = pid * BLOCK_SIZE
offsets = block_start + tl.arange(0, BLOCK_SIZE)
mask = offsets < n_elements
# Load x and gate from global memory (single read each)
x = tl.load(x_ptr + offsets, mask=mask)
gate = tl.load(gate_ptr + offsets, mask=mask)
# Compute SiLU in registers: silu(x) = x * sigmoid(x)
sigmoid_x = 1.0 / (1.0 + tl.exp(-x))
silu_x = x * sigmoid_x
# Fused multiply with gate (in registers)
output = silu_x * gate
# Single write to global memory
tl.store(output_ptr + offsets, output, mask=mask)
def fast_silu_multiply(x: torch.Tensor, gate: torch.Tensor) -> torch.Tensor:
\"\"\"Drop-in replacement for slow_silu_multiply using the fused Triton kernel.\"\"\"
output = torch.empty_like(x)
n_elements = x.numel()
BLOCK_SIZE = 1024
grid = (triton.cdiv(n_elements, BLOCK_SIZE),)
fused_silu_multiply_kernel[grid](x, gate, output, n_elements, BLOCK_SIZE=BLOCK_SIZE)
return output
if __name__ == '__main__':
x = torch.randn(4096, device='cuda')
gate = torch.randn(4096, device='cuda')
out = fast_silu_multiply(x, gate)
print(f'Output shape: {out.shape}, mean: {out.mean().item():.4f}')
print('Kernel executed successfully.')
"""
},
{"action_type": "run_script", "filename": "fast_silu_kernel.py"},
{"action_type": "submit"},
],
}
# ---------------------------------------------------------------------------
# Agent loop
# ---------------------------------------------------------------------------
def run_task_with_llm(client: openai.OpenAI, task_id: str, verbose: bool = True) -> float:
"""Run an LLM agent against the given task using the OpenAI client. Returns final grader score."""
print(f"[START] task={task_id}", flush=True)
if verbose:
print(f"\n{'='*60}")
print(f" TASK: {task_id}")
print(f"{'='*60}")
# Reset env
reset_resp = env_reset(task_id)
obs = reset_resp["observation"]
# --- SAFELY GRAB THE README CONTENT ---
files_dict = obs.get('files', {})
file_keys = list(files_dict.keys())
# Extract just the first part of the task ID (e.g., 'task1' from 'task1_security_audit')
task_short = task_id.split("_")[0]
readme_key = f"README_{task_short}.txt"
# Safely get the content. If the README doesn't exist, it defaults to an empty string
readme_content = files_dict.get(readme_key, "")
# --------------------------------------
messages = [
{"role": "system", "content": SYSTEM_PROMPT},
{"role": "user", "content": (
f"Task: {obs['message']}\n\n"
f"Available files: {file_keys}\n\n"
f"README:\n{readme_content}\n\n"
"Begin solving. Output your first action as a JSON object."
)}
]
actual_steps = 0
for step_num in range(MAX_STEPS_PER_TASK):
actual_steps += 1
if verbose:
print(f"\n Step {step_num + 1}/{MAX_STEPS_PER_TASK}:")
# Call LLM with retry
max_retries = 3
raw = ""
for attempt in range(max_retries):
try:
response = client.chat.completions.create(
model=MODEL_NAME,
messages=messages,
temperature=0.1,
max_tokens=2000,
)
# FIX: Access the first choice in the choices list
if response.choices and response.choices[0].message.content:
raw = response.choices[0].message.content.strip()
break
except Exception as e:
if verbose:
print(f" API error (attempt {attempt+1}/{max_retries}): {e}")
if attempt == max_retries - 1:
print(f"[END] task={task_id} score=0.0010 steps={actual_steps}", flush=True)
return 0.001
time.sleep(3)
if not raw:
if verbose:
print(" Failed to get valid text from API response. Skipping step.")
messages.append({"role": "user", "content": "Your last response was empty or blocked. Please provide a valid JSON action."})
continue
if verbose:
print(f" Agent: {raw[:200]}")
# Parse JSON action
try:
start = raw.find("{")
end = raw.rfind("}") + 1
if start >= 0 and end > start:
action = json.loads(raw[start:end])
else:
if verbose:
print(" Could not parse JSON action. Skipping.")
continue
except json.JSONDecodeError as e:
if verbose:
print(f" JSON parse error: {e}")
continue
# Send action to env
try:
step_resp = env_step(action)
except Exception as e:
if verbose:
print(f" Env step error: {e}")
break
obs = step_resp["observation"]
reward = step_resp["reward"]["value"]
done = step_resp["done"]
expl = step_resp["reward"]["explanation"]
print(f"[STEP] step={actual_steps} reward={reward:.4f}", flush=True)
if verbose:
print(f" Reward: {reward:+.3f} | {expl[:100]}")
print(f" Partial score: {obs['partial_score']:.3f}")
messages.append({"role": "assistant", "content": raw})
warning = ""
if reward < 0:
warning = "WARNING: Your last action resulted in a negative reward or failure. DO NOT repeat the exact same action. Try a different approach.\n\n"
messages.append({"role": "user", "content": (
f"Step result (Step {step_num + 1}/{MAX_STEPS_PER_TASK}):\n"
f"Reward: {reward:.4f}\n"
f"Explanation: {expl}\n"
f"Metrics: {json.dumps(obs['metrics'])}\n"
f"Files now on filesystem: {list(obs['files'].keys())}\n\n"
f"{warning}"
f"{'Episode done.' if done else 'Continue solving. Output your next action as EXACTLY ONE JSON object.'}"
)})
if done:
break
time.sleep(0.2)
grade_resp = env_grader()
score = grade_resp["score"]
print(f"[END] task={task_id} score={score:.4f} steps={actual_steps}", flush=True)
if verbose:
print(f"\n Final grader score: {score:.4f}")
print(f" Passed: {grade_resp['passed']}")
return score
def run_task_with_expert(task_id: str, verbose: bool = True) -> float:
"""Run the deterministic expert solution. Used when no API key is set."""
print(f"[START] task={task_id}", flush=True)
if verbose:
print(f"\n{'='*60}")
print(f" TASK: {task_id} (expert solution β no API key)")
print(f"{'='*60}")
env_reset(task_id)
actions = EXPERT_SOLUTIONS.get(task_id, [])
actual_steps = 0
for i, action in enumerate(actions):
actual_steps += 1
if verbose:
print(f" Step {i+1}: {action['action_type']} {action.get('filename', '')}")
resp = env_step(action)
reward = resp["reward"]["value"]
expl = resp["reward"]["explanation"]
print(f"[STEP] step={actual_steps} reward={reward:.4f}", flush=True)
if verbose:
print(f" Reward: {reward:+.3f} | {expl[:100]}")
if resp["done"]:
break
time.sleep(0.1)
grade_resp = env_grader()
score = grade_resp["score"]
print(f"[END] task={task_id} score={score:.4f} steps={actual_steps}", flush=True)
if verbose:
print(f"\n Final grader score: {score:.4f} | Passed: {grade_resp['passed']}")
return score
# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------
def main():
parser = argparse.ArgumentParser(description="FrontierLabs-Env Baseline Agent")
parser.add_argument("--task", type=str, default=None, help="Run single task by ID")
parser.add_argument("--json-output", action="store_true", help="Output JSON for /baseline endpoint")
parser.add_argument("--url", type=str, default=None, help="Override environment URL")
args = parser.parse_args()
global BASE_URL
if args.url:
BASE_URL = args.url
tasks_to_run = (
[args.task] if args.task
else ["task1_security_audit", "task2_fsdp_cluster", "task3_triton_kernel"]
)
use_llm = bool(HF_TOKEN)
client = openai.OpenAI(base_url=API_BASE_URL, api_key=HF_TOKEN) if use_llm else None
verbose = not args.json_output
if verbose:
print("FrontierLabs-Env Baseline Agent")
print(f"Model: {MODEL_NAME if use_llm else 'expert (no HF_TOKEN)'}")
print(f"Server: {BASE_URL}")
results = {}
for task_id in tasks_to_run:
try:
if use_llm and client:
score = run_task_with_llm(client, task_id, verbose=verbose)
else:
score = run_task_with_expert(task_id, verbose=verbose)
results[task_id] = {"score": round(score, 4), "passed": score >= 0.8}
except Exception as e:
import traceback
print(f"\n[CRASH DETECTED ON {task_id}]")
traceback.print_exc()
# π΄ CHANGE: Set fallback score to 0.001 instead of 0.0
print(f"[END] task={task_id} score=0.0010 steps=0", flush=True)
results[task_id] = {"score": 0.001, "passed": False, "error": str(e)}
if verbose:
print(f" ERROR on {task_id}: {e}")
# π΄ CHANGE: Set fallback average to 0.001
avg = sum(r["score"] for r in results.values()) / len(results) if results else 0.001
summary = {
"model": MODEL_NAME if use_llm else "expert",
"task_results": results,
"average_score": round(avg, 4),
"all_passed": all(r["passed"] for r in results.values()),
}
if verbose:
print(f"\n{'='*60}")
print(" BASELINE RESULTS SUMMARY")
print(f"{'='*60}")
for tid, res in results.items():
status = "β
PASS" if res["passed"] else "β FAIL"
print(f" {tid}: {res['score']:.4f} {status}")
print(f"\n Average Score: {avg:.4f}")
print(f" All Tasks Passed: {summary['all_passed']}")
else:
# JSON output for /baseline endpoint parsing
print(json.dumps(summary))
return summary
if __name__ == "__main__":
main() |