synth-ai 0.2.14__py3-none-any.whl → 0.4.4__py3-none-any.whl

synth_ai/environments/examples/tictactoe/taskset.py

@@ -1,215 +0,0 @@
-from __future__ import annotations
-
-import random
-from dataclasses import dataclass
-from typing import List
-from uuid import uuid4
-
-import numpy as np
-
-from synth_ai.environments.tasks.core import (
-    Impetus,
-    Intent,
-    SplitInfo,
-    TaskInstance,
-    TaskInstanceMetadata,
-    TaskInstanceSet,
-)
-
-from .engine import COORD_TO_IDX, PLAYER_MARKS, WIN_PATTERNS
-
-
-@dataclass
-class TicTacToeTaskInstanceMetadata(TaskInstanceMetadata):
-    starting_player: str  # "X" or "O"
-    opening_moves: List[str]  # Pre-made moves to create position
-    optimal_outcome: str  # "win", "draw", "loss" for starting player
-    position_complexity: int  # Number of pre-moves made
-    shortest_win_length: int  # Min moves to force win/draw
-
-
-@dataclass
-class TicTacToeTaskInstance(TaskInstance):
-    async def serialize(self) -> dict:
-        return {
-            "id": str(self.id),
-            "impetus": {"instructions": self.impetus.instructions},
-            "intent": {
-                "rubric": self.intent.rubric,
-                "gold_trajectories": self.intent.gold_trajectories,
-                "gold_state_diff": self.intent.gold_state_diff,
-            },
-            "metadata": {
-                "starting_player": self.metadata.starting_player,
-                "opening_moves": self.metadata.opening_moves,
-                "optimal_outcome": self.metadata.optimal_outcome,
-                "position_complexity": self.metadata.position_complexity,
-                "shortest_win_length": self.metadata.shortest_win_length,
-            },
-            "is_reproducible": self.is_reproducible,
-            "initial_engine_snapshot": self.initial_engine_snapshot,
-        }
-
-    @classmethod
-    async def deserialize(cls, data: dict) -> "TicTacToeTaskInstance":
-        from uuid import UUID
-
-        metadata = TicTacToeTaskInstanceMetadata(
-            starting_player=data["metadata"]["starting_player"],
-            opening_moves=data["metadata"]["opening_moves"],
-            optimal_outcome=data["metadata"]["optimal_outcome"],
-            position_complexity=data["metadata"]["position_complexity"],
-            shortest_win_length=data["metadata"]["shortest_win_length"],
-        )
-
-        return cls(
-            id=UUID(data["id"]),
-            impetus=Impetus(instructions=data["impetus"]["instructions"]),
-            intent=Intent(
-                rubric=data["intent"]["rubric"],
-                gold_trajectories=data["intent"]["gold_trajectories"],
-                gold_state_diff=data["intent"]["gold_state_diff"],
-            ),
-            metadata=metadata,
-            is_reproducible=data["is_reproducible"],
-            initial_engine_snapshot=data["initial_engine_snapshot"],
-        )
-
-
-def _evaluate_position(board: np.ndarray, player: int) -> str:
-    """Simple evaluation of position outcome with perfect play"""
-    # Check for immediate win
-    for pattern in WIN_PATTERNS:
-        values = [board[i] for i in pattern]
-        if values.count(player) == 3:
-            return "win"
-        if values.count(3 - player) == 3:
-            return "loss"
-
-    # Check if board is full
-    if np.all(board != 0):
-        return "draw"
-
-    # For simplicity, assume draw for non-terminal positions
-    # In a real implementation, this would use minimax
-    return "draw"
-
-
-def _count_shortest_win(board: np.ndarray, player: int) -> int:
-    """Count minimum moves to force a win/draw"""
-    # Simplified: return remaining empty cells
-    empty_cells = sum(1 for i in range(9) if board[i] == 0)
-    return max(1, empty_cells // 2)
-
-
-async def create_tictactoe_taskset() -> TaskInstanceSet:
-    """Generate diverse TicTacToe starting positions"""
-    instances = []
-
-    # Configuration for different position types
-    POSITION_CONFIGS = {
-        "opening": {"pre_moves": 0, "count": 10},  # Fresh games
-        "early": {"pre_moves": 1, "count": 15},  # After 1 move
-        "mid": {"pre_moves": 2, "count": 15},  # After 2 moves
-        "complex": {"pre_moves": 3, "count": 10},  # After 3 moves
-    }
-
-    all_coords = list(COORD_TO_IDX.keys())
-
-    for config_name, config in POSITION_CONFIGS.items():
-        for i in range(config["count"]):
-            # Generate random opening moves
-            opening_moves = []
-            board = np.zeros(9, dtype=int)
-            current_player = "X"
-
-            # Make pre-moves
-            available_coords = all_coords.copy()
-            for move_idx in range(config["pre_moves"]):
-                if not available_coords:
-                    break
-
-                # Random move
-                move = random.choice(available_coords)
-                opening_moves.append(move)
-                available_coords.remove(move)
-
-                # Update board
-                board[COORD_TO_IDX[move]] = PLAYER_MARKS[current_player]
-                current_player = "O" if current_player == "X" else "X"
-
-            # Evaluate position
-            starting_player = current_player
-            optimal_outcome = _evaluate_position(board, PLAYER_MARKS[starting_player])
-            shortest_win = _count_shortest_win(board, PLAYER_MARKS[starting_player])
-
-            # Create metadata
-            metadata = TicTacToeTaskInstanceMetadata(
-                starting_player=starting_player,
-                opening_moves=opening_moves,
-                optimal_outcome=optimal_outcome,
-                position_complexity=config["pre_moves"],
-                shortest_win_length=shortest_win,
-            )
-
-            # Create instance
-            impetus = Impetus(
-                instructions=(
-                    f"You are playing TicTacToe as {starting_player}. "
-                    + "The game is played on a 3x3 grid with cells labeled A1-A3, B1-B3, C1-C3. "
-                    + (
-                        f"The game has already had {len(opening_moves)} moves."
-                        if opening_moves
-                        else "This is a fresh game."
-                    )
-                    + f" You must place your mark ({starting_player}) in an empty cell. "
-                    + "Win by getting three of your marks in a row (horizontally, vertically, or diagonally)."
-                )
-            )
-
-            intent = Intent(
-                rubric={"goal": f"Win the game as {starting_player}, or at least force a draw"},
-                gold_trajectories=None,
-                gold_state_diff={"optimal_outcome": optimal_outcome},
-            )
-
-            instance = TicTacToeTaskInstance(
-                id=uuid4(),
-                impetus=impetus,
-                intent=intent,
-                metadata=metadata,
-                is_reproducible=True,
-                initial_engine_snapshot=None,
-            )
-
-            instances.append(instance)
-
-    # Shuffle instances
-    random.shuffle(instances)
-
-    # Define splits based on complexity
-    val_ids = {inst.id for inst in instances if inst.metadata.position_complexity == 1}
-    test_ids = {inst.id for inst in instances if inst.metadata.position_complexity >= 2}
-
-    # If not enough instances for splits, use simple division
-    if len(val_ids) == 0 or len(test_ids) == 0:
-        total = len(instances)
-        val_end = int(total * 0.15)
-        test_end = int(total * 0.30)
-        val_ids = {instances[i].id for i in range(val_end)}
-        test_ids = {instances[i].id for i in range(val_end, test_end)}
-
-    split_info = SplitInfo(
-        val_instance_ids=val_ids, test_instance_ids=test_ids, _is_split_defined=True
-    )
-
-    return TaskInstanceSet(
-        name="TicTacToe Procedural TaskSet",
-        description="Procedurally generated TicTacToe tasks with varying starting positions.",
-        instances=instances,
-        split_info=split_info,
-    )
-
-
-# Make taskset available as module attribute
-taskset = create_tictactoe_taskset

synth_ai/environments/examples/verilog/__init__.py

@@ -1,10 +0,0 @@
-from .engine import VerilogEngine
-from .environment import VerilogEnvironment
-from .taskset import VerilogTaskInstance, create_verilog_taskset
-
-__all__ = [
-    "VerilogEngine",
-    "VerilogEnvironment",
-    "VerilogTaskInstance",
-    "create_verilog_taskset",
-]

synth_ai/environments/examples/verilog/engine.py

@@ -1,421 +0,0 @@
-from __future__ import annotations
-
-import shutil
-import subprocess
-from dataclasses import dataclass
-from pathlib import Path
-from typing import Any, Dict, Optional, Tuple
-
-from synth_ai.environments.environment.rewards.core import RewardComponent, RewardStack
-from synth_ai.environments.stateful.engine import StatefulEngine, StatefulEngineSnapshot
-from synth_ai.environments.tasks.core import TaskInstance
-
-
-@dataclass
-class VerilogEngineSnapshot(StatefulEngineSnapshot):
-    task_instance_dict: Dict
-    engine_snapshot: Dict
-
-    def model_dump(self) -> Dict:
-        """Convert dataclass to dictionary for compatibility with Pydantic models."""
-        return {
-            "task_instance_dict": self.task_instance_dict,
-            "engine_snapshot": self.engine_snapshot,
-        }
-
-
-@dataclass
-class VerilogPublicState:
-    files: Dict[str, str]
-    build_dir: str
-    task_completed: bool = False
-    last_compile_output: Optional[str] = None
-    last_simulate_output: Optional[str] = None
-
-
-@dataclass
-class VerilogPrivateState:
-    reward_last: float
-    total_reward: float
-    terminated: bool
-    truncated: bool
-
-
-class VerilogCompileSuccessComponent(RewardComponent):
-    async def score(self, state: VerilogPublicState, action: Any) -> float:
-        if hasattr(action, "get") and action.get("type") == "compile":
-            # Check if compilation was successful (returncode 0)
-            if action.get("returncode") == 0:
-                return 0.01  # Normalized: 0.1 / 10.0 = 0.01
-        return 0.0
-
-
-class VerilogSimulationPassComponent(RewardComponent):
-    async def score(self, state: VerilogPublicState, action: Any) -> float:
-        if hasattr(action, "get") and action.get("type") == "simulate":
-            # Check if simulation passed
-            if action.get("passed", False):
-                return 0.1  # Normalized: 1.0 / 10.0 = 0.1
-        return 0.0
-
-
-class VerilogStepPenaltyComponent(RewardComponent):
-    def __init__(self, penalty: float = 0.0):  # No per-step reward - only reward accomplishments
-        self.penalty = penalty
-
-    async def score(self, state: Any, action: Any) -> float:
-        return self.penalty
-
-
-class VerilogSubmitSuccessComponent(RewardComponent):
-    """Reward for successful submission (tests passed). Max reward = 1.0 (normalized)."""
-    async def score(self, state: VerilogPublicState, action: Any) -> float:
-        if hasattr(action, "get") and action.get("type") == "submit":
-            # Check if submission passed
-            if action.get("passed", False):
-                return 1.0  # Normalized: Maximum reward is now 1.0
-        return 0.0
-
-
-class VerilogEngine(StatefulEngine):
-    """
-    Stateful Verilog evaluation engine with persistent artifact snapshots.
-    """
-
-    def __init__(self, task_instance: TaskInstance):
-        # Validate required Verilog tools are available
-        self._validate_verilog_tools()
-        
-        self.task_instance = task_instance
-        self._total_reward = 0.0
-        self._current_action_for_reward: Optional[Dict[str, Any]] = None
-
-        self.reward_stack = RewardStack(
-            components=[
-                VerilogCompileSuccessComponent(),
-                VerilogSimulationPassComponent(),
-                VerilogSubmitSuccessComponent(),
-                VerilogStepPenaltyComponent(penalty=0.0),  # No per-step reward
-            ]
-        )
-
-        # Initialize paths - will be set properly in _reset_engine
-        self.snapshot_dir: Optional[Path] = None
-        self.build_dir: Optional[Path] = None
-
-        # Track last compile/simulate outputs
-        self._last_compile_output: Optional[str] = None
-        self._last_simulate_output: Optional[str] = None
-    
-    @staticmethod
-    def _validate_verilog_tools() -> None:
-        """Validate that required Verilog tools (iverilog, vvp) are available."""
-        missing_tools = []
-        
-        if not shutil.which("iverilog"):
-            missing_tools.append("iverilog")
-        if not shutil.which("vvp"):
-            missing_tools.append("vvp")
-        
-        if missing_tools:
-            error_msg = (
-                f"🚨🚨🚨 CRITICAL CONFIGURATION ERROR 🚨🚨🚨\n"
-                f"\n"
-                f"Missing required Verilog tools: {', '.join(missing_tools)}\n"
-                f"\n"
-                f"The Verilog environment CANNOT function without these tools.\n"
-                f"ALL compile/simulate operations will FAIL.\n"
-                f"ALL rewards will be ZERO.\n"
-                f"Training or evaluation will be COMPLETELY BROKEN.\n"
-                f"\n"
-                f"🔧 FIX THIS NOW:\n"
-                f"1. Add 'iverilog' to apt_packages in Modal deployment config\n"
-                f"2. Location: examples/task_apps/verilog/task_app/grpo_verilog.py\n"
-                f"3. Look for: modal=ModalDeploymentConfig(\n"
-                f"4. Add: apt_packages=('iverilog',)  # Provides both iverilog and vvp\n"
-                f"5. Redeploy: uvx synth-ai modal-serve grpo-verilog\n"
-                f"\n"
-                f"{'='*80}"
-            )
-            print(f"\n{'='*80}\n{error_msg}\n{'='*80}\n", flush=True)
-            raise RuntimeError(error_msg)
-
-    async def _reset_engine(
-        self, *, seed: Optional[int] = None
-    ) -> Tuple[VerilogPrivateState, VerilogPublicState]:
-        """Initialize the Verilog environment with task files."""
-        self._total_reward = 0.0
-        self._current_action_for_reward = None
-        self._last_compile_output = None
-        self._last_simulate_output = None
-
-        # Initialize snapshot from task instance
-        self._init_snapshot()
-
-        priv = VerilogPrivateState(
-            reward_last=0.0, total_reward=0.0, terminated=False, truncated=False
-        )
-
-        pub = VerilogPublicState(
-            files=self._get_file_contents(),
-            build_dir=str(self.build_dir),
-            task_completed=False,
-        )
-
-        return priv, pub
-
-    async def _step_engine(
-        self, action_result: Dict[str, Any]
-    ) -> Tuple[VerilogPrivateState, VerilogPublicState]:
-        """Process an action result and update engine state."""
-        self._current_action_for_reward = action_result
-        
-        # DEBUG: Print action_result
-        print(f"\n[ENGINE DEBUG] _step_engine called")
-        print(f"  action_result: {action_result}")
-        print(f"  action_result.type: {action_result.get('type')}")
-        print(f"  action_result.returncode: {action_result.get('returncode')}")
-        print(f"  action_result.ok: {action_result.get('ok')}")
-
-        # Update last outputs if this is a compile or simulate action
-        if action_result.get("type") == "compile":
-            stdout = action_result.get("stdout", "")
-            stderr = action_result.get("stderr", "")
-            # Combine stdout and stderr for compile output, stderr has the error info
-            self._last_compile_output = stderr if stderr else stdout
-        elif action_result.get("type") == "simulate":
-            self._last_simulate_output = action_result.get("stdout")
-
-        # Calculate reward using RewardStack
-        current_pub_state = VerilogPublicState(
-            files=self._get_file_contents(),
-            build_dir=str(self.build_dir),
-            task_completed=action_result.get("submitted", False) and action_result.get("passed", False),
-        )
-
-        reward_from_stack = await self.reward_stack.step_reward(
-            state=current_pub_state, action=self._current_action_for_reward
-        )
-        self._current_action_for_reward = None
-        
-        # DEBUG: Print reward
-        print(f"[ENGINE DEBUG] reward_from_stack: {reward_from_stack}")
-
-        self._total_reward += reward_from_stack
-
-        # Check termination conditions - only terminate if submitted (regardless of pass/fail)
-        terminated = action_result.get("submitted", False)
-
-        priv = VerilogPrivateState(
-            reward_last=reward_from_stack,
-            total_reward=self._total_reward,
-            terminated=terminated,
-            truncated=False,
-        )
-
-        pub = VerilogPublicState(
-            files=self._get_file_contents(),
-            build_dir=str(self.build_dir),
-            task_completed=action_result.get("submitted", False) and action_result.get("passed", False),
-            last_compile_output=self._last_compile_output,
-            last_simulate_output=self._last_simulate_output,
-        )
-
-        return priv, pub
-
-    def _init_snapshot(self) -> None:
-        """Initialize snapshot directory from task instance data."""
-        if not hasattr(self.task_instance, "snapshot_dir"):
-            raise ValueError("Task instance must have a snapshot_dir attribute")
-
-        self.snapshot_dir = Path(self.task_instance.snapshot_dir)
-
-        if self.snapshot_dir.exists() and any(self.snapshot_dir.iterdir()):
-            # Already initialized
-            self.build_dir = self.snapshot_dir / "build"
-            self.build_dir.mkdir(exist_ok=True)
-            return
-
-        # Copy pristine files from task data
-        pristine_dir = getattr(self.task_instance, "pristine_dir", None)
-        if pristine_dir and Path(pristine_dir).exists():
-            shutil.copytree(pristine_dir, self.snapshot_dir, dirs_exist_ok=True)
-        else:
-            # Create basic structure if no pristine dir
-            self.snapshot_dir.mkdir(parents=True, exist_ok=True)
-
-        self.build_dir = self.snapshot_dir / "build"
-        self.build_dir.mkdir(exist_ok=True)
-
-    def _get_file_contents(self) -> Dict[str, str]:
-        """Get contents of all Verilog files in the snapshot directory."""
-        if not self.snapshot_dir:
-            return {}
-
-        files = {}
-        for p in self.snapshot_dir.rglob("*.v"):
-            try:
-                relative_path = p.relative_to(self.snapshot_dir)
-                files[str(relative_path)] = p.read_text()
-            except Exception:
-                continue
-        return files
-
-    async def write_file(self, path: str, content: str) -> Dict[str, Any]:
-        """Write content to a file in the snapshot directory."""
-        if not self.snapshot_dir:
-            return {"ok": False, "error": "Snapshot directory not initialized"}
-
-        file_path = self.snapshot_dir / path
-        file_path.parent.mkdir(parents=True, exist_ok=True)
-        file_path.write_text(content)
-        return {"ok": True, "type": "write_file"}
-
-    async def compile(
-        self, sources: Optional[list] = None, testbench: Optional[str] = None
-    ) -> Dict[str, Any]:
-        """Compile Verilog sources with iverilog."""
-        if not self.snapshot_dir or not self.build_dir:
-            return {"ok": False, "error": "Directories not initialized"}
-
-        # Default to all .v files if no sources specified
-        if sources is None:
-            sources = [str(p.relative_to(self.snapshot_dir)) for p in self.snapshot_dir.glob("*.v")]
-
-        src_paths = [self.snapshot_dir / src for src in sources]
-
-        # Add testbench if specified
-        if testbench:
-            tb_path = self.snapshot_dir / testbench
-            if tb_path.exists():
-                src_paths.append(tb_path)
-
-        binary = self.build_dir / "a.out"
-        cmd = ["iverilog", "-g2012", "-o", str(binary)] + [str(p) for p in src_paths]
-
-        try:
-            proc = subprocess.run(cmd, capture_output=True, text=True, timeout=30)
-            return {
-                "ok": proc.returncode == 0,
-                "type": "compile",
-                "stdout": proc.stdout,
-                "stderr": proc.stderr,
-                "returncode": proc.returncode,
-                "binary": str(binary) if proc.returncode == 0 else None,
-            }
-        except subprocess.TimeoutExpired:
-            return {"ok": False, "error": "Compilation timeout", "type": "compile"}
-        except FileNotFoundError:
-            error_msg = (
-                "🚨 CRITICAL ERROR: 'iverilog' executable not found! 🚨\n"
-                "The Verilog compiler (iverilog) is not installed in this environment.\n"
-                "This will cause ALL compile operations to fail and result in ZERO rewards.\n"
-                "Fix: Add 'iverilog' to apt_packages in the Modal deployment config.\n"
-                "Location: examples/task_apps/verilog/task_app/grpo_verilog.py -> modal=ModalDeploymentConfig(apt_packages=('iverilog',))"
-            )
-            print(f"\n{'='*80}\n{error_msg}\n{'='*80}\n", flush=True)
-            raise RuntimeError(error_msg) from None
-        except Exception as e:
-            return {"ok": False, "error": str(e), "type": "compile"}
-
-    async def simulate(self, binary: Optional[str] = None) -> Dict[str, Any]:
-        """Run vvp on compiled binary."""
-        if not self.build_dir:
-            return {"ok": False, "error": "Build directory not initialized"}
-
-        bin_path = binary if binary else str(self.build_dir / "a.out")
-
-        try:
-            proc = subprocess.run(["vvp", bin_path], capture_output=True, text=True, timeout=30)
-
-            # Check for various success indicators
-            stdout = proc.stdout
-            passed = (
-                "ALL_TESTS_PASSED" in stdout
-                or ("Mismatches: 0 " in stdout and "samples" in stdout)
-                or ("no mismatches" in stdout.lower() and "errors" not in stdout.lower())
-            )
-
-            return {
-                "ok": True,
-                "type": "simulate",
-                "stdout": proc.stdout,
-                "stderr": proc.stderr,
-                "returncode": proc.returncode,
-                "passed": passed,
-            }
-        except subprocess.TimeoutExpired:
-            return {"ok": False, "error": "Simulation timeout", "type": "simulate"}
-        except FileNotFoundError:
-            error_msg = (
-                "🚨 CRITICAL ERROR: 'vvp' executable not found! 🚨\n"
-                "The Verilog simulator (vvp) is not installed in this environment.\n"
-                "This will cause ALL simulate operations to fail and result in ZERO rewards.\n"
-                "Fix: Add 'iverilog' to apt_packages in the Modal deployment config (provides both iverilog and vvp).\n"
-                "Location: examples/task_apps/verilog/task_app/grpo_verilog.py -> modal=ModalDeploymentConfig(apt_packages=('iverilog',))"
-            )
-            print(f"\n{'='*80}\n{error_msg}\n{'='*80}\n", flush=True)
-            raise RuntimeError(error_msg) from None
-        except Exception as e:
-            return {"ok": False, "error": str(e), "type": "simulate"}
-
-    async def submit(self) -> Dict[str, Any]:
-        """Submit solution for grading."""
-        # Check if the last simulation passed
-        # Parse the last simulation output to determine if tests passed
-        passed = False
-        detail = "No simulation run yet"
-        
-        if self._last_simulate_output:
-            stdout = self._last_simulate_output
-            passed = (
-                "ALL_TESTS_PASSED" in stdout
-                or ("Mismatches: 0 " in stdout and "samples" in stdout)
-                or ("no mismatches" in stdout.lower() and "errors" not in stdout.lower())
-            )
-            if passed:
-                detail = "All tests passed"
-            else:
-                detail = "Tests failed - please review simulation output"
-        
-        return {
-            "ok": True,
-            "type": "submit",
-            "passed": passed,
-            "detail": detail,
-            "submitted": True,
-        }
-
-    async def _serialize_engine(self) -> VerilogEngineSnapshot:
-        """Serialize engine state to a snapshot."""
-        engine_data = {
-            "total_reward": self._total_reward,
-            "snapshot_dir": str(self.snapshot_dir) if self.snapshot_dir else None,
-            "build_dir": str(self.build_dir) if self.build_dir else None,
-        }
-
-        task_instance_dict = await self.task_instance.serialize()
-
-        return VerilogEngineSnapshot(
-            task_instance_dict=task_instance_dict, engine_snapshot=engine_data
-        )
-
-    @classmethod
-    async def _deserialize_engine(cls, snapshot: VerilogEngineSnapshot) -> "VerilogEngine":
-        """Deserialize engine from snapshot."""
-        # This would need proper task instance deserialization
-        # For now, create a minimal implementation
-        from synth_ai.environments.examples.verilog.taskset import VerilogTaskInstance
-
-        task_instance = await VerilogTaskInstance.deserialize(snapshot.task_instance_dict)
-        engine = cls(task_instance)
-
-        engine_data = snapshot.engine_snapshot
-        engine._total_reward = engine_data.get("total_reward", 0.0)
-
-        if engine_data.get("snapshot_dir"):
-            engine.snapshot_dir = Path(engine_data["snapshot_dir"])
-        if engine_data.get("build_dir"):
-            engine.build_dir = Path(engine_data["build_dir"])
-
-        return engine