synth-ai 0.2.13.dev2__py3-none-any.whl → 0.2.16__py3-none-any.whl

synth_ai/demos/math/task_app_entry.py

@@ -0,0 +1,51 @@
+"""Task app registry entry for the math demo Modal deployment."""
+
+from __future__ import annotations
+
+from contextlib import suppress
+from importlib import import_module
+
+from synth_ai.task.apps import ModalDeploymentConfig, TaskAppEntry, register_task_app
+
+try:
+    from synth_ai.task.apps.math_single_step import build_config as base_build_config
+except ModuleNotFoundError:
+    base_module = import_module("examples.rl.task_app.math_single_step")
+    base_build_config = base_module.build_config
+
+DEMO_MODAL_CONFIG = ModalDeploymentConfig(
+    app_name="hendrycks-math-task-app",
+    pip_packages=(
+        "fastapi>=0.110.0",
+        "uvicorn>=0.23.0",
+        "pydantic>=2.6.0",
+        "httpx>=0.24.0",
+        "numpy>=1.24.0",
+        "aiohttp>=3.8.0",
+        "datasets>=2.16.0",
+        "synth-ai",
+    ),
+)
+
+
+def build_config():
+    """Reuse the shared math single-step TaskAppConfig."""
+
+    return base_build_config()
+
+
+def register_demo_entry() -> None:
+    entry = TaskAppEntry(
+        app_id="hendrycks-math-demo",
+        description="Demo math task app (Modal-focused) shipping with synth-ai demos.",
+        config_factory=build_config,
+        modal=DEMO_MODAL_CONFIG,
+    )
+    with suppress(ValueError):
+        register_task_app(entry=entry)
+
+
+register_demo_entry()
+
+
+__all__ = ["DEMO_MODAL_CONFIG", "build_config", "register_demo_entry"]

synth_ai/environments/environment/core.py

@@ -1,4 +1,10 @@
-from synth_ai.core.system import System
+class System:
+    """Minimal base data structure shared by environment types."""
+
+    id: str
+    name: str
+    description: str
+    pass
 
 
 class Environment(System):

synth_ai/environments/examples/bandit/engine.py

@@ -4,7 +4,6 @@ from dataclasses import dataclass
 from typing import Any
 
 import numpy as np
-
 from synth_ai.environments.environment.shared_engine import (
     GetObservationCallable,
     InternalObservation,

synth_ai/environments/examples/bandit/environment.py

@@ -3,7 +3,6 @@ from __future__ import annotations
 from typing import Any
 
 from pydantic import BaseModel, Field, ValidationError
-
 from synth_ai.environments.environment.shared_engine import (
     GetObservationCallable,
     InternalObservation,

synth_ai/environments/examples/crafter_classic/environment.py

@@ -397,7 +397,7 @@ class CrafterClassicEnvironment(StatefulEnvironment, ReproducibleEnvironment[Cra
             priv_state, pub_state, self.custom_step_observation_callable
         )
         total_step_time = time.time() - step_start_time
-        logger.info(
+        logger.debug(
             f"CrafterClassic step completed in {total_step_time:.3f}s (interact: {interact_time:.3f}s)"
         )
         return obs

synth_ai/environments/examples/verilog/engine.py

@@ -46,7 +46,7 @@ class VerilogCompileSuccessComponent(RewardComponent):
         if hasattr(action, "get") and action.get("type") == "compile":
             # Check if compilation was successful (returncode 0)
             if action.get("returncode") == 0:
-                return 0.1
+                return 0.01  # Normalized: 0.1 / 10.0 = 0.01
         return 0.0
 
 
@@ -55,12 +55,12 @@ class VerilogSimulationPassComponent(RewardComponent):
         if hasattr(action, "get") and action.get("type") == "simulate":
             # Check if simulation passed
             if action.get("passed", False):
-                return 1.0
+                return 0.1  # Normalized: 1.0 / 10.0 = 0.1
         return 0.0
 
 
 class VerilogStepPenaltyComponent(RewardComponent):
-    def __init__(self, penalty: float = -0.01):
+    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:
@@ -68,12 +68,12 @@ class VerilogStepPenaltyComponent(RewardComponent):
 
 
 class VerilogSubmitSuccessComponent(RewardComponent):
-    """Reward for successful submission (tests passed)."""
+    """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 10.0  # Large reward for completing the task correctly
+                return 1.0  # Normalized: Maximum reward is now 1.0
         return 0.0
 
 
@@ -83,6 +83,9 @@ class VerilogEngine(StatefulEngine):
     """
 
     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
@@ -92,7 +95,7 @@ class VerilogEngine(StatefulEngine):
                 VerilogCompileSuccessComponent(),
                 VerilogSimulationPassComponent(),
                 VerilogSubmitSuccessComponent(),
-                VerilogStepPenaltyComponent(penalty=-0.01),
+                VerilogStepPenaltyComponent(penalty=0.0),  # No per-step reward
             ]
         )
 
@@ -103,6 +106,39 @@ class VerilogEngine(StatefulEngine):
         # 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
@@ -133,6 +169,13 @@ class VerilogEngine(StatefulEngine):
     ) -> 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":
@@ -147,18 +190,21 @@ class VerilogEngine(StatefulEngine):
         current_pub_state = VerilogPublicState(
             files=self._get_file_contents(),
             build_dir=str(self.build_dir),
-            task_completed=action_result.get("passed", False),
+            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
-        terminated = action_result.get("passed", False) or action_result.get("submitted", False)
+        # Check termination conditions - only terminate if submitted (regardless of pass/fail)
+        terminated = action_result.get("submitted", False)
 
         priv = VerilogPrivateState(
             reward_last=reward_from_stack,
@@ -170,7 +216,7 @@ class VerilogEngine(StatefulEngine):
         pub = VerilogPublicState(
             files=self._get_file_contents(),
             build_dir=str(self.build_dir),
-            task_completed=action_result.get("passed", False),
+            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,
         )
@@ -259,6 +305,16 @@ class VerilogEngine(StatefulEngine):
             }
         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"}
 
@@ -290,6 +346,16 @@ class VerilogEngine(StatefulEngine):
             }
         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"}
 

synth_ai/environments/examples/wordle/environment.py

@@ -3,7 +3,6 @@ from __future__ import annotations
 from typing import Any
 
 from pydantic import BaseModel, Field
-
 from synth_ai.environments.environment.shared_engine import (
     GetObservationCallable,
     InternalObservation,

synth_ai/evals/base.py

@@ -1,13 +1,24 @@
+from __future__ import annotations
+
+from abc import ABC, abstractmethod
+from typing import Any
+
+
 class Judgement:
     def __init__(
-        self, criteria: str, score: float, reasoning: str = "", evidence: list[str] = None
-    ):
+        self,
+        criteria: str,
+        score: float,
+        reasoning: str = "",
+        evidence: list[str] | None = None,
+    ) -> None:
         self.criteria = criteria
         self.score = score
         self.reasoning = reasoning
         self.evidence = evidence or []
 
 
-class BaseEval:
-    async def run(self, data: any) -> list[Judgement]:
-        pass
+class BaseEval(ABC):
+    @abstractmethod
+    async def run(self, data: Any) -> list[Judgement]:
+        """Execute the evaluation and return a list of judgements."""

synth_ai/evals/client.py

@@ -10,7 +10,7 @@ import os
 import warnings
 from typing import Any, Literal, TypedDict
 
-from synth_ai.http import AsyncHttpClient, HTTPError
+from synth_ai.http_client import AsyncHttpClient, HTTPError
 from synth_ai.tracing_v3.serialization import normalize_for_json
 
 Provider = Literal["groq", "gemini"]

synth_ai/inference/client.py

@@ -7,7 +7,7 @@ from synth_ai.api.models.supported import (
     normalize_model_identifier,
 )
 
-from ..http import AsyncHttpClient
+from .._utils.http import AsyncHttpClient
 
 
 class InferenceClient:

synth_ai/learning/client.py

@@ -11,7 +11,7 @@ from synth_ai.api.models.supported import (
 )
 from synth_ai.learning.sft.config import prepare_sft_job_payload
 
-from ..http import AsyncHttpClient, HTTPError, sleep
+from .._utils.http import AsyncHttpClient, HTTPError, sleep
 
 
 class LearningClient:

synth_ai/learning/health.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 
 from typing import Any
 
-from ..http import AsyncHttpClient
+from .._utils.http import AsyncHttpClient
 
 
 def _api_base(b: str) -> str:

synth_ai/learning/jobs.py

@@ -5,7 +5,7 @@ from collections.abc import Callable
 from contextlib import suppress
 from typing import Any
 
-from ..http import AsyncHttpClient, sleep
+from .._utils.http import AsyncHttpClient, sleep
 from .constants import TERMINAL_EVENT_FAILURE, TERMINAL_EVENT_SUCCESS, TERMINAL_STATUSES
 
 

synth_ai/learning/rl/client.py

@@ -10,7 +10,7 @@ from synth_ai.api.models.supported import (
     normalize_model_identifier,
 )
 
-from ...http import AsyncHttpClient, HTTPError, sleep
+from ..._utils.http import AsyncHttpClient, HTTPError, sleep
 
 
 def _api_base(b: str) -> str:

synth_ai/learning/rl/env_keys.py

@@ -1,4 +1,4 @@
-"""Helpers for uploading RL environment credentials to the backend."""
+"""Helpers for uploading Environment credentials to the backend."""
 
 from __future__ import annotations
 

synth_ai/learning/rl/secrets.py

@@ -1,4 +1,4 @@
-"""Helpers for generating RL environment credentials."""
+"""Helpers for generating Environment credentials."""
 
 from __future__ import annotations
 

synth_ai/learning/sft/client.py

@@ -3,7 +3,7 @@ from __future__ import annotations
 from pathlib import Path
 from typing import Any
 
-from ...http import AsyncHttpClient, HTTPError
+from ..._utils.http import AsyncHttpClient, HTTPError
 from .config import prepare_sft_job_payload
 from .data import validate_jsonl_or_raise