save

Signed-off-by: Guyue Huang <[email protected]>

Author: guyueh1

examples/configs/distillation_math.yaml

@@ -170,6 +170,7 @@ policy: &POLICY_BASE
         top_k: null
         stop_token_ids: null
         stop_strings: null
+        ignore_eos: false
         vllm_cfg:
             async_engine: false
             precision: ${...precision}

examples/configs/evals/eval.yaml

@@ -16,6 +16,7 @@ generation:
   model_name: "Qwen/Qwen2.5-Math-1.5B-Instruct"
   stop_token_ids: null
   stop_strings: null
+  ignore_eos: false
   vllm_cfg:
     async_engine: false
     precision: "bfloat16"

examples/configs/grpo_math_1B.yaml

@@ -214,6 +214,7 @@ policy:
     top_k: null
     stop_token_ids: null
     stop_strings: null
+    ignore_eos: false
     vllm_cfg:
       async_engine: false
       precision: ${policy.precision}

examples/configs/vlm_grpo_3B.yaml

@@ -203,6 +203,7 @@ policy:
     top_k: null
     stop_token_ids: null
     stop_strings: null
+    ignore_eos: false
     vllm_cfg:
       async_engine: false # Only for internal testing, will be enabled by https://github.com/NVIDIA/NeMo-RL/issues/447.
       precision: ${policy.precision}

examples/configs/vlm_grpo_3B_megatron.yaml

@@ -103,6 +103,7 @@ policy:
     top_k: null
     stop_token_ids: null
     stop_strings: null
+    ignore_eos: false
     vllm_cfg:
       async_engine: false
       precision: ${policy.precision}

examples/run_eval_random_dataset.py

@@ -0,0 +1,135 @@
+# Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import argparse
+import os
+import pprint
+import sys
+
+sys.path.append(os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
+
+from omegaconf import OmegaConf
+from transformers import AutoTokenizer, PreTrainedTokenizerBase
+
+from nemo_rl.algorithms.utils import get_tokenizer
+from nemo_rl.data.datasets import AllTaskProcessedDataset, RandomDataset
+from nemo_rl.distributed.ray_actor_environment_registry import get_actor_python_env
+from nemo_rl.distributed.virtual_cluster import init_ray
+from nemo_rl.environments.dummy_environment import DummyEnvironment
+from nemo_rl.evals.eval import MasterConfig, run_env_eval, setup
+from nemo_rl.models.generation import configure_generation_config
+from nemo_rl.utils.config import load_config, parse_hydra_overrides
+
+TokenizerType = PreTrainedTokenizerBase
+
+
+def parse_args():
+    """Parse command line arguments."""
+    parser = argparse.ArgumentParser(description="Run Evaluation with configuration")
+    parser.add_argument(
+        "--config", type=str, default=None, help="Path to YAML config file"
+    )
+
+    # Parse known args for the script
+    args, overrides = parser.parse_known_args()
+
+    return args, overrides
+
+
+def setup_data(tokenizer: AutoTokenizer, data_config, env_configs):
+    print("Setting up data...")
+
+    # load dataset
+    base_dataset = RandomDataset(data_config["input_len_or_input_len_generator"])
+
+    env = DummyEnvironment.options(
+        runtime_env={
+            "py_executable": get_actor_python_env(
+                "nemo_rl.environments.math_environment.MathEnvironment"
+            )
+        }
+    ).remote()
+
+    dataset = AllTaskProcessedDataset(
+        dataset=base_dataset.formatted_ds["train"],
+        tokenizer=tokenizer,
+        default_task_data_spec=base_dataset.task_spec,
+        task_data_processors=base_dataset.processor,
+        max_seq_length=data_config["max_input_seq_length"],
+    )
+
+    return dataset, env, tokenizer
+
+
+def main():
+    """Main entry point."""
+    # Parse arguments
+    args, overrides = parse_args()
+
+    if not args.config:
+        args.config = os.path.join(
+            os.path.dirname(__file__), "configs", "evals", "eval.yaml"
+        )
+
+    config = load_config(args.config)
+    print(f"Loaded configuration from: {args.config}")
+
+    if overrides:
+        print(f"Overrides: {overrides}")
+        config = parse_hydra_overrides(config, overrides)
+
+    config: MasterConfig = OmegaConf.to_container(config, resolve=True)
+    print("Applied CLI overrides")
+
+    # Print config
+    print("Final config:")
+    pprint.pprint(config)
+
+    # Init ray
+    init_ray()
+
+    # Setup tokenizer
+    tokenizer = get_tokenizer(config["tokenizer"])
+    config["generation"] = configure_generation_config(
+        config["generation"], tokenizer, is_eval=True
+    )
+    config["generation"]["vllm_cfg"]["load_format"] = (
+        "dummy"  # for random dataset eval, we use dummy weight initialization
+    )
+
+    # Setup data
+    (
+        dataset,
+        env,
+        tokenizer,
+    ) = setup_data(tokenizer, config["data"], config["env"])
+
+    # Setup
+    (
+        vllm_generation,
+        dataloader,
+        master_config,
+    ) = setup(config, tokenizer, dataset)
+
+    # Run evaluation
+    run_env_eval(
+        vllm_generation,
+        dataloader,
+        env,
+        master_config,
+    )
+
+
+if __name__ == "__main__":
+    main()