fix(gym): make the gym venv prefetch dry run pass again

examples/nemo_gym/prefetch_super_all_envs.yaml

@@ -60,6 +60,8 @@ env:
 
     # Thin vllm_model clients for the judge refs the bundled configs expect (so refs resolve and
     # the vllm_model venv builds). genrm_model is already defined by the bundled genrm_compare.yaml.
+    # return_token_id_information/uses_reasoning_parser have no defaults in Gym's VLLMModelConfig,
+    # so they must be set even on thin clients that never spin up a server.
     safety_judge_model:
       responses_api_models:
         vllm_model:
@@ -68,6 +70,8 @@ env:
           api_key: dummy
           model: dummy-judge
           spinup_server: false
+          return_token_id_information: false
+          uses_reasoning_parser: false
     judge_model:
       responses_api_models:
         vllm_model:
@@ -76,6 +80,8 @@ env:
           api_key: dummy
           model: dummy-judge
           spinup_server: false
+          return_token_id_information: false
+          uses_reasoning_parser: false
     # Thin client for the *_reasoning_off agent refs.
     policy_model_reasoning_off:
       responses_api_models:
@@ -85,17 +91,24 @@ env:
           api_key: dummy
           model: dummy-model
           spinup_server: false
+          return_token_id_information: false
+          uses_reasoning_parser: false
     # Standalone server defined only to bake the local_vllm_model venv (the runtime GPU judges
     # use this type). dry_run builds the venv without loading a model.
+    # LocalVLLMModelConfig extends VLLMModelConfig, so it needs the two fields above plus a
+    # vllm_serve_env_vars mapping (also default-less).
     prefetch_local_vllm_model:
       responses_api_models:
         local_vllm_model:
           entrypoint: app.py
           model: dummy-model
+          return_token_id_information: false
+          uses_reasoning_parser: false
           vllm_serve_kwargs:
             tensor_parallel_size: 1
             data_parallel_size: 1
             pipeline_parallel_size: 1
+          vllm_serve_env_vars: {}
 
     # swe_agents marks container_formatter/dataset_path mandatory; dummies satisfy the dry run.
     swe_agents_train:

examples/nemo_gym/prefetch_ultra_all_envs.yaml

@@ -82,6 +82,8 @@ env:
 
     # Thin vllm_model clients for the judge refs the bundled configs expect (so refs resolve and
     # the vllm_model venv builds). genrm_model is already defined by the bundled genrm_compare.yaml.
+    # return_token_id_information/uses_reasoning_parser have no defaults in Gym's VLLMModelConfig,
+    # so they must be set even on thin clients that never spin up a server.
     safety_judge_model:
       responses_api_models:
         vllm_model:
@@ -90,6 +92,8 @@ env:
           api_key: dummy
           model: dummy-judge
           spinup_server: false
+          return_token_id_information: false
+          uses_reasoning_parser: false
     judge_model:
       responses_api_models:
         vllm_model:
@@ -98,6 +102,8 @@ env:
           api_key: dummy
           model: dummy-judge
           spinup_server: false
+          return_token_id_information: false
+          uses_reasoning_parser: false
     # Thin client for the *_reasoning_off agent refs.
     policy_model_reasoning_off:
       responses_api_models:
@@ -107,17 +113,24 @@ env:
           api_key: dummy
           model: dummy-model
           spinup_server: false
+          return_token_id_information: false
+          uses_reasoning_parser: false
     # Standalone server defined only to bake the local_vllm_model venv (the runtime GPU judges
     # use this type). dry_run builds the venv without loading a model.
+    # LocalVLLMModelConfig extends VLLMModelConfig, so it needs the two fields above plus a
+    # vllm_serve_env_vars mapping (also default-less).
     prefetch_local_vllm_model:
       responses_api_models:
         local_vllm_model:
           entrypoint: app.py
           model: dummy-model
+          return_token_id_information: false
+          uses_reasoning_parser: false
           vllm_serve_kwargs:
             tensor_parallel_size: 1
             data_parallel_size: 1
             pipeline_parallel_size: 1
+          vllm_serve_env_vars: {}
 
     # swe_agents marks container_formatter/dataset_path mandatory; dummies satisfy the dry run.
     swe_agents_train:

examples/nemo_gym/prefetch_venvs.py

@@ -98,6 +98,15 @@ def prefetch_nemo_gym_venvs(config_paths: list[str]) -> None:
                         **os.environ,
                         "VIRTUAL_ENV": nemo_gym_py_exec,
                         "UV_PROJECT_ENVIRONMENT": nemo_gym_py_exec,
+                        # Gym server setup scripts shell out to `uv pip install --no-cache`
+                        # (e.g. swe_agents' r2e_gym.sh), and uv rejects --no-cache combined
+                        # with symlink installs: "Symlink-based installation is not supported
+                        # with `--no-cache`". docker/Dockerfile exports UV_LINK_MODE=symlink
+                        # for this prefetch, and os.environ above propagates it into every
+                        # server's setup subprocess. Pin hardlink for the Gym actor only, so
+                        # the outer prefetch keeps symlink semantics. hardlink is equally
+                        # space-efficient and, unlike symlink, survives uv cache pruning.
+                        "UV_LINK_MODE": "hardlink",
                     },
                 },
             }