openenv/repl

1# Copyright (c) Meta Platforms, Inc. and affiliates.2# All rights reserved.3#4# This source code is licensed under the BSD-style license found in the5# LICENSE file in the root directory of this source tree.6 7"""8FastAPI application for the REPL Environment.9 10This module creates an HTTP server that exposes the REPLEnvironment11over HTTP and WebSocket endpoints, compatible with EnvClient.12 13The server includes llm_query and llm_query_batched support via HuggingFace Inference API,14enabling the Recursive Language Model (RLM) paradigm.15 16LLM Token Configuration:17    1. Client can pass `hf_token` in reset() - RECOMMENDED18    2. Server fallback: HF_TOKEN environment variable19 20LLM functions are created dynamically in REPLEnvironment.reset() based on the21available token (client or server).22 23Usage:24    # Development (with auto-reload):25    uvicorn server.app:app --reload --host 0.0.0.0 --port 800026 27    # Production:28    uvicorn server.app:app --host 0.0.0.0 --port 8000 --workers 429 30    # Or run directly:31    uv run --project . server32 33Environment Variables:34    HF_TOKEN: Fallback HuggingFace API token (client token takes priority)35    LLM_MODEL: Model to use for llm_query/llm_query_batched (default: Qwen/Qwen3.5-9B)36"""37 38import inspect39import logging40import os41 42try:43    from openenv.core.env_server.http_server import create_app44 45    from ..models import REPLAction, REPLObservation46    from .gradio_ui import build_repl_gradio_app47    from .repl_environment import REPLEnvironment48except ImportError:49    from models import REPLAction, REPLObservation50    from openenv.core.env_server.http_server import create_app51    from server.gradio_ui import build_repl_gradio_app52    from server.repl_environment import REPLEnvironment53 54 55# ============== CONFIGURATION ==============56LLM_MODEL = os.environ.get("LLM_MODEL", "Qwen/Qwen3.5-9B")57HF_TOKEN = os.environ.get("HF_TOKEN")58REPL_MAX_ITERATIONS = int(os.environ.get("REPL_MAX_ITERATIONS", "30"))59REPL_MAX_OUTPUT_LENGTH = int(os.environ.get("REPL_MAX_OUTPUT_LENGTH", "8192"))60REPL_CONTEXT_PREVIEW_LENGTH = int(os.environ.get("REPL_CONTEXT_PREVIEW_LENGTH", "500"))61REPL_RLM_MAX_DEPTH = int(os.environ.get("REPL_RLM_MAX_DEPTH", "2"))62REPL_RLM_MAX_ITERATIONS = int(os.environ.get("REPL_RLM_MAX_ITERATIONS", "30"))63# ==========================================64 65_logger = logging.getLogger(__name__)66 67# Log LLM configuration68if HF_TOKEN:69    print("[REPL Server] LLM support ENABLED (server token configured)")70    print(f"[REPL Server] Default model: {LLM_MODEL}")71else:72    print("[REPL Server] No server HF_TOKEN configured")73    print(74        "[REPL Server] LLM functions will be enabled if client passes hf_token in reset()"75    )76 77 78def create_repl_environment() -> REPLEnvironment:79    """Factory function that creates REPLEnvironment with server config.80 81    LLM functions are created dynamically during `reset()` when a client82    passes `hf_token`. Rewards are computed via the default `REPLRubric`;83    pass `expected_answer` at reset time for outcome-based scoring.84    """85    return REPLEnvironment(86        max_iterations=REPL_MAX_ITERATIONS,87        max_output_length=REPL_MAX_OUTPUT_LENGTH,88        context_preview_length=REPL_CONTEXT_PREVIEW_LENGTH,89        rlm_max_depth=REPL_RLM_MAX_DEPTH,90        rlm_max_iterations=REPL_RLM_MAX_ITERATIONS,91    )92 93 94# Create the app with web interface and README integration.95_sig = inspect.signature(create_app)96if "gradio_builder" in _sig.parameters:97    app = create_app(98        create_repl_environment,99        REPLAction,100        REPLObservation,101        env_name="repl_env",102        max_concurrent_envs=8,103        gradio_builder=build_repl_gradio_app,104    )105else:106    _logger.warning(107        "Installed openenv-core does not support gradio_builder; "108        "custom REPL Gradio tab will not be available."109    )110    app = create_app(111        create_repl_environment,112        REPLAction,113        REPLObservation,114        env_name="repl_env",115        max_concurrent_envs=8,116    )117 118 119def main():120    """121    Entry point for direct execution via uv run or python -m.122 123    This function enables running the server without Docker:124        uv run --project . server125        python -m envs.repl_env.server.app126        openenv serve repl_env127    """128    import uvicorn129 130    uvicorn.run(app, host="0.0.0.0", port=8000)131 132 133if __name__ == "__main__":134    main()135