Spaces:
Sleeping
Sleeping
File size: 16,067 Bytes
08aa12c | 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 | # Copyright (c) Meta Platforms, Inc. and affiliates.
# All rights reserved.
#
# This source code is licensed under the BSD-style license found in the
# LICENSE file in the root directory of this source tree.
"""
Rtl Debugger Environment Implementation.
On reset(), a random task is loaded from the tasks/ directory.
Each task folder must contain:
design_buggy.v - the immutable original buggy module shown to the agent
makefile - the cocotb makefile (must reference design_active.v)
test_*.py - the cocotb testbench(es) that produce result.json
design_active.v is created on reset() as a mutable working copy of design_buggy.v.
The agent overwrites design_active.v each step, so it iterates on its own previous fix.
"""
import difflib
import json
import os
import random
import re
import shutil
import subprocess
import traceback
from uuid import uuid4
from openenv.core.env_server.interfaces import Environment
from openenv.core.env_server.types import State
try:
from ..models import EditOp, RtlDebuggerAction, RtlDebuggerObservation
except ImportError:
from models import EditOp, RtlDebuggerAction, RtlDebuggerObservation
from .graders import get_grader
# Tasks directory is usually at the project root.
# We first try the current working directory (e.g. /app/env in Docker),
# then fall back to the package-relative path.
_TASKS_DIR = os.path.join(os.getcwd(), "tasks")
if not os.path.isdir(_TASKS_DIR):
_TASKS_DIR = os.path.abspath(os.path.join(os.path.dirname(__file__), "..", "tasks"))
def _list_tasks() -> list[str]:
"""Return sorted list of task directory paths found under _TASKS_DIR.
A valid task must contain design_buggy.v and a makefile.
"""
if not os.path.isdir(_TASKS_DIR):
return []
return sorted(
os.path.join(_TASKS_DIR, name)
for name in os.listdir(_TASKS_DIR)
if os.path.isdir(os.path.join(_TASKS_DIR, name))
and os.path.exists(os.path.join(_TASKS_DIR, name, "design_buggy.v"))
and os.path.exists(os.path.join(_TASKS_DIR, name, "makefile"))
)
def _read(path: str) -> str:
with open(path, "r") as f:
return f.read()
def _levenshtein_line_distance(original: str, modified: str) -> int:
"""
Compute a line-based edit distance between two strings.
Treats line addition, deletion, and substitution as 1 operation each.
"""
import difflib
orig_lines = original.splitlines()
mod_lines = modified.splitlines()
matcher = difflib.SequenceMatcher(None, orig_lines, mod_lines)
distance = 0
for tag, i1, i2, j1, j2 in matcher.get_opcodes():
if tag == 'replace':
# substitution of a block of lines
distance += max(i2 - i1, j2 - j1)
elif tag == 'insert':
distance += (j2 - j1)
elif tag == 'delete':
distance += (i2 - i1)
return distance
def _numbered_code(file_path: str) -> str:
"""Return the contents of a file with 1-indexed line numbers prepended."""
lines = _read(file_path).splitlines()
return "\n".join(f"{i + 1:4d}| {line}" for i, line in enumerate(lines))
def _apply_edits(file_path: str, edits: list[EditOp]) -> str:
"""Apply a list of EditOp operations to a file and write back.
Line numbers in each EditOp refer to the file state BEFORE any edits
in this batch. Operations are applied in reverse line-number order so
that earlier indices stay valid as later lines are modified.
Supports multi-line edits via ``end_line`` (range) and
newlines inside ``new_content``.
Returns the patched file content (also written to *file_path*).
"""
lines = _read(file_path).splitlines()
# Sort by line_number descending so later edits don't shift earlier indices.
sorted_edits = sorted(edits, key=lambda e: e.line_number, reverse=True)
for edit in sorted_edits:
start = edit.line_number - 1 # 0-indexed start
end = (edit.end_line or edit.line_number) - 1 # 0-indexed end (inclusive)
end = min(end, len(lines) - 1) # clamp to file length
if edit.op == "replace":
if 0 <= start < len(lines):
new_lines = edit.new_content.split("\n")
lines[start:end + 1] = new_lines
elif edit.op == "insert_after":
insert_pos = max(start + 1, 0) if edit.line_number > 0 else 0
new_lines = edit.new_content.split("\n")
for i, nl in enumerate(new_lines):
lines.insert(insert_pos + i, nl)
elif edit.op == "delete":
if 0 <= start < len(lines):
del lines[start:end + 1]
patched = "\n".join(lines) + "\n"
with open(file_path, "w") as f:
f.write(patched)
return patched
class RtlDebuggerEnvironment(Environment):
"""
RTL Debugging environment.
On reset(), a random task is selected from the tasks/ directory.
- design_buggy.v is read and shown to agent as the original reference.
- design_active.v is created as a fresh mutable copy of design_buggy.v.
On each step(), the agent submits revised code which:
1. Is written to design_active.v (overwrites the previous attempt).
2. Is compiled with iverilog — fatal gate if it fails.
3. Is simulated via `make` (cocotb + icarus).
4. Has result.json parsed for pass/fail metrics.
5. Levenshtein distance is measured from the *previous* design_active.v.
A separate grader() function computes the final normalized score [0,1].
"""
SUPPORTS_CONCURRENT_SESSIONS: bool = True
def __init__(self):
"""Initialize the rtl_debugger environment."""
self._state = State(episode_id=str(uuid4()), step_count=0)
self._task_dir: str = ""
self._task_context: str = "" # contents of context.md
self._task_name: str = ""
@property
def _active_design_path(self) -> str:
return os.path.join(self._task_dir, "design_active.v")
def reset(self, options: dict | None = None) -> RtlDebuggerObservation: # type: ignore[override]
"""
Reset the environment and load a task.
Runs the initial 'buggy' design through simulation to provide baseline feedback.
"""
self._state = State(episode_id=str(uuid4()), step_count=0)
tasks = _list_tasks()
if not tasks:
return RtlDebuggerObservation(
task_id="",
feedback=f"No tasks found in {_TASKS_DIR}.",
compiled=False,
passed_tests=False,
done=False,
reward=0.0,
)
# Handle explicit task selection via options
selected_task_dir = None
if options and "TASK_NAME" in options:
target_name = options["TASK_NAME"]
for t_dir in tasks:
if os.path.basename(t_dir) == target_name:
selected_task_dir = t_dir
break
if not selected_task_dir:
selected_task_dir = random.choice(tasks)
self._task_dir = selected_task_dir
self._task_name = os.path.basename(self._task_dir)
buggy_path = os.path.join(self._task_dir, "design_buggy.v")
context_path = os.path.join(self._task_dir, "context.md")
if os.path.exists(context_path):
self._task_context = _read(context_path)
else:
self._task_context = ""
# Initialise design_active.v from the buggy design
shutil.copy(buggy_path, self._active_design_path)
result_path = os.path.join(self._task_dir, "result.json")
if os.path.exists(result_path):
os.remove(result_path)
# Run baseline simulation to provide initial feedback
baseline_obs = self._run_evaluation(lev_distance=0)
# Override baseline feedback to include a header
baseline_obs.feedback = f"--- (Buggy code feedback) ---\n" + baseline_obs.feedback
return baseline_obs
def step(self, action: RtlDebuggerAction) -> RtlDebuggerObservation: # type: ignore[override]
"""
Apply the agent's line-edit operations, then compile and simulate.
"""
self._state.step_count += 1
try:
# --- Snapshot previous state for Levenshtein ---
prev_active = _read(self._active_design_path)
# --- Apply edits to design_active.v ---
patched = _apply_edits(self._active_design_path, action.edits)
lev_distance = _levenshtein_line_distance(prev_active, patched)
return self._run_evaluation(lev_distance=lev_distance)
except Exception as exc:
traceback.print_exc()
return RtlDebuggerObservation(
task_id=self._task_name,
numbered_code=_numbered_code(self._active_design_path),
feedback=f"[Internal Server Error] {type(exc).__name__}: {exc}",
compiled=False,
passed_tests=False,
pass_rate=0.0,
progress_ratio=0,
reward=-1.0,
done=False,
)
def _run_evaluation(self, lev_distance: int) -> RtlDebuggerObservation:
"""
Internal helper to compile, simulate, and score the current design_active.v.
Used by both reset() and step().
"""
# --- 2. Run simulation via make ---
# Note: 'make' handles both compilation and simulation via cocotb.
try:
make_res = subprocess.run(
["make"],
cwd=self._task_dir,
capture_output=True,
text=True,
timeout=60,
)
make_feedback = make_res.stdout + make_res.stderr
except subprocess.TimeoutExpired:
return RtlDebuggerObservation(
task_id=self._task_name,
numbered_code=_numbered_code(self._active_design_path),
task_context=self._task_context,
feedback="[Simulation Timeout] Your design likely contains an infinite loop. Simulation killed after 60s.",
compiled=False,
passed_tests=False,
pass_rate=0.0,
progress_ratio=0,
levenshtein_distance=lev_distance,
reward=-5.0,
done=False,
)
# --- 3. Parse result.json (written by cocotb into the task dir) ---
result_path = os.path.join(self._task_dir, "result.json")
if not os.path.exists(result_path):
error_lines = []
for line in make_feedback.splitlines():
lower_line = line.lower()
if any(kw in lower_line for kw in ["error", "exception", "fatal", "traceback", "assert"]):
error_lines.append(line)
# If no explicit error keyword found, just keep the tail of the log
if not error_lines:
error_lines = make_feedback.splitlines()[-15:]
filtered_feedback = "\n".join(error_lines)
return RtlDebuggerObservation(
task_id=self._task_name,
numbered_code=_numbered_code(self._active_design_path),
task_context=self._task_context,
feedback=f"=== Simulation/Build Error ===\nSimulation crashed before producing results.\nError Log:\n{filtered_feedback}",
compiled=False,
passed_tests=False,
pass_rate=0.0,
progress_ratio=0,
levenshtein_distance=lev_distance,
reward=-3.0,
done=False,
)
with open(result_path, "r") as f:
result_data = json.load(f)
num_tests = result_data.get("num_tests", 1)
num_passed = result_data.get("num_passed", 0)
pass_rate = num_passed / num_tests if num_tests > 0 else 0.0
passed_all = result_data.get("passed", False)
# Base reward for overall correctness
reward = pass_rate * 12.0
# Bonus reward for sequential correctness (fixing earliest bugs first)
results_list = result_data.get("results", [])
sequential_passes = 0
for r in results_list:
if r.get("pass", True):
sequential_passes += 1
else:
break
progress_ratio = sequential_passes / max(num_tests, 1)
reward += progress_ratio * 3.0
reward -= min(lev_distance * 0.01, 1.0)
if lev_distance == 0 and self._state.step_count > 0:
reward -= 1.5
if passed_all:
reward += 20.0
seq_count = result_data.get("sequence_correctness")
trans_count = result_data.get("transition_correctness")
if seq_count and trans_count:
feedback_lines = [
f"Compiled: OK | Tests: {num_passed}/{num_tests} passed | Seq: {seq_count}/{num_tests} | Trans: {trans_count}/{num_tests} | First failure after: {progress_ratio:.2f} of sequence"
]
else:
feedback_lines = [
f"Compiled: OK | Tests: {num_passed}/{num_tests} passed | First failure after: {progress_ratio:.2f} of sequence"
]
# Parse failed test cases from result.json
if "results" in result_data:
failed_tests = [r for r in result_data["results"] if not r.get("pass", True)]
if failed_tests:
feedback_lines.append("\n=== Failed Test Cases ===")
for ft in failed_tests[:10]:
inp = ft.get("inputs") or ft.get("input", {})
inputs_str = ", ".join(f"{k}={v}" for k, v in inp.items()) if isinstance(inp, dict) else str(inp)
cycle = ft.get("test_id", "?")
expected = ft.get("expected_output", "?")
actual = ft.get("actual_output", "?")
state_info = ""
if "transition_to" in ft:
was_s = ft.get("actual_state_meaning", f"S{ft.get('actual_state', '?')}")
went_s = ft.get("transition_to_meaning", f"S{ft.get('transition_to', '?')}")
exp_s = ft.get("expected_next_state_meaning", f"S{ft.get('expected_next_state', '?')}")
state_info = f" | was {was_s}, went to {went_s} (expected {exp_s})"
feedback_lines.append(f"- Cycle {cycle}: [{inputs_str}] -> Expected: {expected}, Got: {actual}{state_info}")
feedback = "\n".join(filter(None, feedback_lines))
grader = get_grader(self._task_name)
final_score = grader(self._state, self)
print(feedback)
return RtlDebuggerObservation(
task_id=self._task_name,
numbered_code=_numbered_code(self._active_design_path),
task_context=self._task_context,
feedback=feedback,
compiled=True,
passed_tests=passed_all,
pass_rate=pass_rate,
progress_ratio=progress_ratio,
levenshtein_distance=lev_distance,
reward=reward,
done=passed_all,
score=final_score,
)
@property
def state(self) -> State:
"""
Get the current environment state.
Returns:
Current State with episode_id and step_count
"""
return self._state
|