fix: ensure that seeds are instantiated in the same way

src/cogelot/entrypoints/interactive_evaluate.py

@@ -6,6 +6,7 @@
 import torch
 import typer
 from loguru import logger
+from pytorch_lightning import seed_everything
 from rich.console import Console
 from rich.table import Table
 
@@ -29,6 +30,8 @@ def create_evaluation_module(config_path: Path) -> EvaluationLightningModule:
         config_path.name,
         overrides=["model.model.wandb_run_id=8lkml12g", "environment@model.environment=display"],
     )
+    seed = config.get("seed")
+    seed_everything(seed, workers=True)
     evaluation = hydra.utils.instantiate(config["model"])
     assert isinstance(evaluation, EvaluationLightningModule)
     return evaluation

src/vima_bench/env/base.py

@@ -3,6 +3,7 @@
 import os
 import tempfile
 import time
+from contextlib import suppress
 from typing import Literal
 
 import gym
@@ -159,6 +160,12 @@ def set_task(self, task: BaseTask | str | None = None, task_kwargs: dict | None
         # setup task
         ALL_TASKS = _ALL_TASKS.copy()
         ALL_TASKS.update({k.split("/")[1]: v for k, v in ALL_TASKS.items()})
+
+        with suppress(AttributeError):
+            updated_task_kwargs = {"seed": self.global_seed[0]}
+            if task_kwargs:
+                task_kwargs = {**task_kwargs, **updated_task_kwargs}
+
         if isinstance(task, str):
             assert task in ALL_TASKS, f"Invalid task name provided {task}"
             task = ALL_TASKS[task](debug=self._debug, **(task_kwargs or {}))