game_reasoning_arena / ui /gradio_config_generator.py
lcipolina's picture
Fixed typo
e578f77 verified
#!/usr/bin/env python3
"""
Gradio Configuration Generator
This module creates configurations compatible with the existing runner.py and
simulate.py infrastructure, eliminating code duplication in the Gradio app.
"""
import tempfile
import yaml
from typing import Dict, Any, Tuple, List, Optional
import logging
logger = logging.getLogger(__name__)
def _legal_actions_with_labels(env, pid: int) -> List[Tuple[int, str]]:
"""Return the current player's legal actions as (id, label) pairs."""
try:
actions = env.state.legal_actions(pid)
except Exception:
return []
labelled = []
for a in actions:
label = None
if hasattr(env, "get_action_display"):
try:
label = env.get_action_display(a, pid)
except Exception:
label = None
elif hasattr(env.state, "action_to_string"):
try:
label = env.state.action_to_string(pid, a)
except Exception:
label = None
labelled.append((a, label or str(a)))
return labelled
def start_game_interactive(
game_name: str,
player1_type: str,
player2_type: str,
player1_model: Optional[str],
player2_model: Optional[str],
rounds: int,
seed: int,
) -> Tuple[str, Dict[str, Any], List[Tuple[int, str]], List[Tuple[int, str]]]:
"""Initialize env + policies; return (log, state, legal_p0, legal_p1)."""
from src.game_reasoning_arena.arena.utils.seeding import set_seed
from src.game_reasoning_arena.backends import initialize_llm_registry
from src.game_reasoning_arena.arena.games.registry import registry
from src.game_reasoning_arena.arena.agents.policy_manager import (
initialize_policies,
)
cfg = create_config_for_gradio_game(
game_name=game_name,
player1_type=player1_type,
player2_type=player2_type,
player1_model=player1_model,
player2_model=player2_model,
rounds=1,
seed=seed,
)
set_seed(seed)
try:
initialize_llm_registry()
except Exception:
# ok if LLM backend not available for random/human vs random/human
pass
# Build agents + env using your existing infra
policies = initialize_policies(cfg, game_name, seed)
env = registry.make_env(game_name, cfg)
obs, _ = env.reset(seed=seed)
# Map policy order to player ids (same as your simulate.py)
player_to_agent: Dict[int, Any] = {}
for i, policy_name in enumerate(policies.keys()):
player_to_agent[i] = policies[policy_name]
log = []
log.append("๐ŸŽฎ INTERACTIVE GAME")
log.append("=" * 50)
log.append(f"Game: {game_name.replace('_', ' ').title()}")
log.append("")
# Choose which agent_id's board to show:
# - If P0 is human -> agent_id=0; elif P1 is human -> agent_id=1; else 0.
show_id = 0 if player1_type == "human" else (1 if player2_type == "human" else 0)
try:
board = env.render_board(show_id)
log.append("Initial board:")
log.append(board)
except NotImplementedError:
log.append("Board rendering not implemented for this game.")
except Exception as e:
log.append(f"Board not available: {e}")
state = {
"env": env,
"obs": obs,
"terminated": False,
"truncated": False,
"rewards": {0: 0, 1: 0},
"players": {
0: {"type": player1_type},
1: {"type": player2_type},
},
"agents": player_to_agent,
"show_id": show_id,
}
# Auto-advance the game until it's a human player's turn
def _is_human(pid: int) -> bool:
return ((pid == 0 and player1_type == "human") or
(pid == 1 and player2_type == "human"))
def _any_human_needs_action() -> bool:
"""Check if any human player needs to make an action."""
try:
if env.state.is_simultaneous_node():
return _is_human(0) or _is_human(1)
else:
cur = env.state.current_player()
return _is_human(cur)
except Exception:
return False
# Process AI moves until a human needs to act or game ends
term = False
trunc = False
while not (term or trunc) and not _any_human_needs_action():
# Build actions for current turn
if env.state.is_simultaneous_node():
actions = {}
# P0
if not _is_human(0):
response = player_to_agent[0](obs[0])
a0, _ = _extract_action_and_reasoning(response)
actions[0] = a0
# P1
if not _is_human(1):
response = player_to_agent[1](obs[1])
a1, _ = _extract_action_and_reasoning(response)
actions[1] = a1
log.append(f"Auto-play: P0={actions.get(0, 'waiting')}, "
f"P1={actions.get(1, 'waiting')}")
else:
# Sequential game
cur = env.state.current_player()
if not _is_human(cur):
response = player_to_agent[cur](obs[cur])
a, reasoning = _extract_action_and_reasoning(response)
actions = {cur: a}
log.append(f"Player {cur} chooses {a}")
if reasoning and reasoning != "None":
prev = reasoning[:100]
if len(reasoning) > 100:
prev += "..."
log.append(f" Reasoning: {prev}")
else:
# Human's turn - break out of loop
break
# Step env
obs, step_rewards, term, trunc, _ = env.step(actions)
for pid, r in step_rewards.items():
state["rewards"][pid] += r
# Update board display
try:
log.append("Board:")
log.append(env.render_board(show_id))
except NotImplementedError:
log.append("Board rendering not implemented for this game.")
except Exception as e:
log.append(f"Board not available: {e}")
# Update state with current observations
state["obs"] = obs
state["terminated"] = term
state["truncated"] = trunc
# Prepare human choices for current state
legal_p0: List[Tuple[int, str]] = []
legal_p1: List[Tuple[int, str]] = []
if not (term or trunc):
try:
if env.state.is_simultaneous_node():
if player1_type == "human":
legal_p0 = _legal_actions_with_labels(env, 0)
if player2_type == "human":
legal_p1 = _legal_actions_with_labels(env, 1)
else:
cur = env.state.current_player()
if cur == 0 and player1_type == "human":
legal_p0 = _legal_actions_with_labels(env, 0)
if cur == 1 and player2_type == "human":
legal_p1 = _legal_actions_with_labels(env, 1)
except Exception:
pass
return "\n".join(log), state, legal_p0, legal_p1
def submit_human_move(
action_p0: Optional[int],
action_p1: Optional[int],
state: Dict[str, Any],
) -> Tuple[str, Dict[str, Any], List[Tuple[int, str]], List[Tuple[int, str]]]:
"""
Process human move and continue advancing the game automatically until:
- It's a human player's turn again, OR
- The game ends
Returns (log_append, state, next_legal_p0, next_legal_p1)
"""
if not state:
return "No game is running.", state, [], []
env = state["env"]
obs = state["obs"]
term = state["terminated"]
trunc = state["truncated"]
rewards = state["rewards"]
ptypes = state["players"]
agents = state["agents"]
show_id = state["show_id"]
if term or trunc:
return "Game already finished.", state, [], []
def _is_human(pid: int) -> bool:
return ptypes[pid]["type"] == "human"
def _any_human_needs_action() -> bool:
"""Check if any human player needs to make an action."""
try:
if env.state.is_simultaneous_node():
return _is_human(0) or _is_human(1)
else:
cur = env.state.current_player()
return _is_human(cur)
except Exception:
return False
log = []
# Continue processing moves until a human needs to act or game ends
while not (term or trunc):
# Build actions for current turn
if env.state.is_simultaneous_node():
actions = {}
# P0
if _is_human(0):
if action_p0 is None:
return ("Pick an action for Player 0.", state,
_legal_actions_with_labels(env, 0), [])
actions[0] = action_p0
action_p0 = None # Only use human action once
else:
a0, _ = _extract_action_and_reasoning(agents[0](obs[0]))
actions[0] = a0
# P1
if _is_human(1):
if action_p1 is None:
return ("Pick an action for Player 1.", state,
[], _legal_actions_with_labels(env, 1))
actions[1] = action_p1
action_p1 = None # Only use human action once
else:
a1, _ = _extract_action_and_reasoning(agents[1](obs[1]))
actions[1] = a1
log.append(f"Actions: P0={actions[0]}, P1={actions[1]}")
else:
# Sequential game
cur = env.state.current_player()
if _is_human(cur):
chosen = action_p0 if cur == 0 else action_p1
if chosen is None:
choices = _legal_actions_with_labels(env, cur)
return ("Pick an action first.", state,
choices if cur == 0 else [],
choices if cur == 1 else [])
actions = {cur: chosen}
log.append(f"Player {cur} (human) chooses {chosen}")
# Clear the action so it's not reused
if cur == 0:
action_p0 = None
else:
action_p1 = None
else:
a, reasoning = _extract_action_and_reasoning(agents[cur](obs[cur]))
actions = {cur: a}
log.append(f"Player {cur} (agent) chooses {a}")
if reasoning and reasoning != "None":
prev = reasoning[:100] + ("..." if len(reasoning) > 100 else "")
log.append(f" Reasoning: {prev}")
# Step env
obs, step_rewards, term, trunc, _ = env.step(actions)
for pid, r in step_rewards.items():
rewards[pid] += r
# Board
try:
log.append("Board:")
log.append(env.render_board(show_id))
except NotImplementedError:
log.append("Board rendering not implemented for this game.")
except Exception as e:
log.append(f"Board not available: {e}")
# Check if game ended
if term or trunc:
break
# Check if we should continue automatically (AI turn) or stop (human turn)
if _any_human_needs_action():
break # Stop here, human needs to act
# If we reach here, it's an AI's turn - continue the loop
# Game ended or waiting for human input
if term or trunc:
if rewards[0] > rewards[1]:
winner = "Player 0"
elif rewards[1] > rewards[0]:
winner = "Player 1"
else:
winner = "Draw"
log.append(f"Winner: {winner}")
log.append(f"Scores: P0={rewards[0]}, P1={rewards[1]}")
state["terminated"] = term
state["truncated"] = trunc
state["obs"] = obs
return "\n".join(log), state, [], []
# Determine next human choices
next_p0, next_p1 = [], []
try:
if env.state.is_simultaneous_node():
if _is_human(0):
next_p0 = _legal_actions_with_labels(env, 0)
if _is_human(1):
next_p1 = _legal_actions_with_labels(env, 1)
else:
cur = env.state.current_player()
if _is_human(cur):
choices = _legal_actions_with_labels(env, cur)
if cur == 0:
next_p0 = choices
else:
next_p1 = choices
except Exception:
pass
state["obs"] = obs
return "\n".join(log), state, next_p0, next_p1
def create_config_for_gradio_game(
game_name: str,
player1_type: str,
player2_type: str,
player1_model: str = None,
player2_model: str = None,
rounds: int = 1,
seed: int = 42,
use_ray: bool = False
) -> Dict[str, Any]:
"""
Create a configuration dictionary compatible with the existing
runner.py and simulate.py infrastructure.
Args:
game_name: Name of the game to play
player1_type: Type of player 1 (human, random, llm)
player2_type: Type of player 2 (human, random, llm)
player1_model: LLM model for player 1 (if applicable)
player2_model: LLM model for player 2 (if applicable)
rounds: Number of episodes to play
seed: Random seed for reproducibility
use_ray: Whether to use Ray for parallel processing
Returns:
Configuration dictionary compatible with runner.py
"""
# Base configuration structure (matches default_simulation_config)
config = {
"env_config": {
"game_name": game_name,
"max_game_rounds": None,
},
"num_episodes": rounds,
"seed": seed,
"use_ray": use_ray,
"mode": f"{player1_type}_vs_{player2_type}",
"agents": {},
"llm_backend": {
"max_tokens": 250,
"temperature": 0.1,
"default_model": "litellm_groq/gemma-7b-it",
},
"log_level": "INFO",
}
# Configure player agents
config["agents"]["player_0"] = _create_agent_config(
player1_type, player1_model)
config["agents"]["player_1"] = _create_agent_config(
player2_type, player2_model)
# Debug: Print the agent configurations
print("๐Ÿ“‹ CONFIG DEBUG: Agent configurations created:")
print(f" Player 0 config: {config['agents']['player_0']}")
print(f" Player 1 config: {config['agents']['player_1']}")
# Update backend default model if LLM is used
# Check player 1 first
if (player1_type == "llm" and player1_model) or player1_type.startswith("llm_"):
if player1_model:
config["llm_backend"]["default_model"] = player1_model
elif player1_type.startswith("llm_"):
# Extract model from player type (e.g., "llm_gpt2" -> "gpt2")
config["llm_backend"]["default_model"] = player1_type[4:]
# Check player 2 if player 1 doesn't have LLM
elif (player2_type == "llm" and player2_model) or player2_type.startswith("llm_"):
if player2_model:
config["llm_backend"]["default_model"] = player2_model
elif player2_type.startswith("llm_"):
# Extract model from player type (e.g., "llm_gpt2" -> "gpt2")
config["llm_backend"]["default_model"] = player2_type[4:]
return config
def _create_agent_config(player_type: str,
model: str = None) -> Dict[str, Any]:
"""
Create agent configuration based on player type and model.
Handles both Gradio-specific formats (e.g., "hf_gpt2", "random_bot")
and standard formats (e.g., "llm", "random", "human").
Args:
player_type: Type of player (human, random, random_bot, hf_*, etc.)
model: Model name for LLM agents
Returns:
Agent configuration dictionary
"""
print("๐Ÿ”ง AGENT CONFIG DEBUG: Creating agent config for:")
print(f" player_type: {player_type}")
print(f" model: {model}")
# Handle Gradio-specific formats
if player_type == "random_bot":
config = {"type": "random"}
elif player_type == "human":
config = {"type": "human"}
elif player_type.startswith("hf_"):
# Extract model from player type (e.g., "hf_gpt2" -> "gpt2")
model_from_type = player_type[3:] # Remove "hf_" prefix
# Use the hf_prefixed model name for LLM registry lookup
model_name = f"hf_{model_from_type}"
config = {
"type": "llm", # Use standard LLM agent type
"model": model_name # This will be looked up in LLM_REGISTRY
}
elif player_type.startswith("llm_"):
# For backwards compatibility with LiteLLM models
model_from_type = player_type[4:] # Remove "llm_" prefix
# Map display model names to actual model names with prefixes
model_name = model or model_from_type
if not model_name.startswith(("litellm_", "vllm_")):
# Add litellm_ prefix for LiteLLM models
model_name = f"litellm_{model_name}"
config = {
"type": "llm",
"model": model_name
}
elif player_type == "llm":
model_name = model or "litellm_groq/gemma-7b-it"
if not model_name.startswith(("litellm_", "vllm_")):
model_name = f"litellm_{model_name}"
config = {
"type": "llm",
"model": model_name
}
elif player_type == "random":
config = {"type": "random"}
else:
# Default to random for unknown types
config = {"type": "random"}
print(f" โ†’ Created config: {config}")
return config
def create_temporary_config_file(config: Dict[str, Any]) -> str:
"""
Create a temporary YAML config file that can be used with runner.py.
Args:
config: Configuration dictionary
Returns:
Path to the temporary config file
"""
# Create temporary file
temp_file = tempfile.NamedTemporaryFile(
mode='w',
suffix='.yaml',
delete=False
)
try:
yaml.dump(config, temp_file, default_flow_style=False)
temp_file.flush()
return temp_file.name
finally:
temp_file.close()
def run_game_with_existing_infrastructure(
game_name: str,
player1_type: str,
player2_type: str,
player1_model: str = None,
player2_model: str = None,
rounds: int = 1,
seed: int = 42
) -> str:
"""
Run a game using the existing runner.py and simulate.py infrastructure,
but capture detailed game logs for Gradio display.
This function reuses the existing simulation infrastructure while providing
detailed game output for the Gradio interface.
Args:
game_name: Name of the game to play
player1_type: Type of player 1
player2_type: Type of player 2
player1_model: LLM model for player 1 (if applicable)
player2_model: LLM model for player 2 (if applicable)
rounds: Number of episodes to play
seed: Random seed
Returns:
Detailed game simulation results as a string
"""
try:
# Import the existing infrastructure
from src.game_reasoning_arena.arena.utils.seeding import set_seed
from src.game_reasoning_arena.backends import initialize_llm_registry
from src.game_reasoning_arena.arena.games.registry import registry
from src.game_reasoning_arena.arena.agents.policy_manager import (
initialize_policies, policy_mapping_fn
)
# Create configuration
config = create_config_for_gradio_game(
game_name=game_name,
player1_type=player1_type,
player2_type=player2_type,
player1_model=player1_model,
player2_model=player2_model,
rounds=rounds,
seed=seed
)
# Set seed
set_seed(seed)
# Initialize LLM registry (required for simulate_game)
initialize_llm_registry()
# Use existing infrastructure but capture detailed logs
return _run_game_with_detailed_logging(game_name, config, seed)
except ImportError as e:
logger.error(f"Failed to import simulation infrastructure: {e}")
return f"Error: Simulation infrastructure not available. {e}"
except Exception as e:
logger.error(f"Game simulation failed: {e}")
return f"Error during game simulation: {e}"
def _run_game_with_detailed_logging(
game_name: str,
config: Dict[str, Any],
seed: int
) -> str:
"""
Run game simulation with detailed logging for Gradio display.
This reuses the existing infrastructure components but captures
detailed game state information for user display.
"""
from src.game_reasoning_arena.arena.games.registry import registry
from src.game_reasoning_arena.arena.agents.policy_manager import (
initialize_policies, policy_mapping_fn
)
# Initialize using existing infrastructure
policies_dict = initialize_policies(config, game_name, seed)
env = registry.make_env(game_name, config)
# Create player mapping (reusing existing logic)
player_to_agent = {}
for i, policy_name in enumerate(policies_dict.keys()):
player_to_agent[i] = policies_dict[policy_name]
game_log = []
# Add header
game_log.append("๐ŸŽฎ GAME SIMULATION RESULTS")
game_log.append("=" * 50)
game_log.append(f"Game: {game_name.replace('_', ' ').title()}")
game_log.append(f"Episodes: {config['num_episodes']}")
game_log.append("")
# Player information
game_log.append("๐Ÿ‘ฅ PLAYERS:")
player1 = config["agents"]["player_0"]
player2 = config["agents"]["player_1"]
game_log.append(f" Player 0: {_format_player_info(player1)}")
game_log.append(f" Player 1: {_format_player_info(player2)}")
game_log.append("")
# Run episodes (reusing compute_actions logic from simulate.py)
for episode in range(config["num_episodes"]):
episode_seed = seed + episode
game_log.append(f"๐ŸŽฏ Episode {episode + 1}")
game_log.append("-" * 30)
observation_dict, _ = env.reset(seed=episode_seed)
terminated = truncated = False
step_count = 0
episode_rewards = {0: 0, 1: 0}
while not (terminated or truncated):
step_count += 1
game_log.append(f"\n๐Ÿ“‹ Step {step_count}")
# Show board state
try:
board = env.render_board(0)
game_log.append("Current board:")
game_log.append(board)
except:
game_log.append("Board state not available")
# Use the existing compute_actions logic from simulate.py
try:
action_dict = _compute_actions_for_gradio(
env, player_to_agent, observation_dict, game_log
)
except Exception as e:
game_log.append(f"โŒ Error computing actions: {e}")
truncated = True
break
# Step forward (reusing existing environment logic)
if not truncated:
observation_dict, rewards, terminated, truncated, _ = env.step(action_dict)
for player_id, reward in rewards.items():
episode_rewards[player_id] += reward
# Episode results
game_log.append(f"\n๐Ÿ Episode {episode + 1} Complete!")
try:
game_log.append("Final board:")
game_log.append(env.render_board(0))
except:
game_log.append("Final board state not available")
if episode_rewards[0] > episode_rewards[1]:
winner = "Player 0"
elif episode_rewards[1] > episode_rewards[0]:
winner = "Player 1"
else:
winner = "Draw"
game_log.append(f"๐Ÿ† Winner: {winner}")
game_log.append(f"๐Ÿ“Š Scores: Player 0={episode_rewards[0]}, Player 1={episode_rewards[1]}")
game_log.append("")
game_log.append("โœ… Simulation completed successfully!")
game_log.append("Check the database logs for detailed move analysis.")
return "\n".join(game_log)
def _compute_actions_for_gradio(env, player_to_agent, observations, game_log):
"""
Compute actions and log details for Gradio display.
This reuses the compute_actions logic from simulate.py.
"""
if env.state.is_simultaneous_node():
# Simultaneous-move game
actions = {}
for player in player_to_agent:
agent_response = player_to_agent[player](observations[player])
action, reasoning = _extract_action_and_reasoning(agent_response)
actions[player] = action
# Always show both action number and action name (universal solution)
try:
action_name = env.state.action_to_string(player, action)
except Exception:
action_name = str(action)
game_log.append(f" Player {player} chooses action {action} ({action_name})")
if reasoning and reasoning != "None":
reasoning_preview = reasoning[:100] + ("..." if len(reasoning) > 100 else "")
game_log.append(f" Reasoning: {reasoning_preview}")
return actions
else:
# Turn-based game
current_player = env.state.current_player()
game_log.append(f"Player {current_player}'s turn")
agent_response = player_to_agent[current_player](observations[current_player])
action, reasoning = _extract_action_and_reasoning(agent_response)
game_log.append(f" Player {current_player} chooses action {action}")
if reasoning and reasoning != "None":
reasoning_preview = reasoning[:100] + ("..." if len(reasoning) > 100 else "")
game_log.append(f" Reasoning: {reasoning_preview}")
return {current_player: action}
def _extract_action_and_reasoning(agent_response):
"""Extract action and reasoning from agent response."""
if isinstance(agent_response, dict) and "action" in agent_response:
action = agent_response.get("action", -1)
reasoning = agent_response.get("reasoning", "None")
return action, reasoning
else:
return agent_response, "None"
def _format_player_info(player_config: Dict[str, Any]) -> str:
"""Format player information for display."""
player_type = player_config["type"]
if player_type == "llm":
model = player_config.get("model", "unknown")
return f"LLM ({model})"
else:
return player_type.replace("_", " ").title()
# For backward compatibility and easy integration
def create_gradio_compatible_config(
game_name: str,
player1_type: str,
player2_type: str,
player1_model: str = None,
player2_model: str = None,
rounds: int = 1
) -> Tuple[Dict[str, Any], str]:
"""
Create both a config dict and a temp file for maximum compatibility.
Returns:
Tuple of (config_dict, temp_file_path)
"""
config = create_config_for_gradio_game(
game_name, player1_type, player2_type,
player1_model, player2_model, rounds
)
temp_file = create_temporary_config_file(config)
return config, temp_file
if __name__ == "__main__":
# Example usage
config = create_config_for_gradio_game(
game_name="tic_tac_toe",
player1_type="llm",
player2_type="random",
player1_model="litellm_groq/llama-3.1-8b-instant",
rounds=3
)
print("Generated configuration:")
print(yaml.dump(config, default_flow_style=False))