NVIDIA-NeMo · wprazuch · Jun 23, 2026 · Jun 23, 2026 · Jun 23, 2026 · Jun 23, 2026
diff --git a/CLAUDE.md b/CLAUDE.md
@@ -82,7 +82,7 @@ uv venv && uv sync --extra dev --group docs
 pre-commit install
 
 # Run servers
-ng_run "+config_paths=[resources_servers/example_single_tool_call/configs/example_single_tool_call.yaml,responses_api_models/vllm_model/configs/vllm_model.yaml]"
+ng_run "+config_paths=[resources_servers/example_single_tool_call/configs/example_single_tool_call.yaml,responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml]"
 
 # Run tests for a specific server (creates .venv per server, installs deps, runs pytest)
 # First run is slow. Use skip_venv_if_present config or place a .venv to skip venv creation.

diff --git a/benchmarks/aalcr/README.md b/benchmarks/aalcr/README.md
@@ -6,7 +6,7 @@ ng_prepare_benchmark "+config_paths=[$config_paths]"
 
 # Run
 ```bash
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/aalcr/config.yaml"
 ng_e2e_collect_rollouts \
     "+config_paths=[${config_paths}]" \

diff --git a/benchmarks/aalcr/config.yaml b/benchmarks/aalcr/config.yaml
@@ -1,7 +1,7 @@
-# Refer to the params in responses_api_models/local_vllm_model/configs/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8.yaml
+# Refer to the params in responses_api_models/vllm_server/configs/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8.yaml
 Qwen3-235B-A22B-Instruct-2507-FP8:
   responses_api_models:
-    vllm_model:
+    vllm_endpoint:
       entrypoint: app.py
       model: Qwen3-235B-A22B-Instruct-2507-FP8
       replace_developer_role_with_system: true

diff --git a/benchmarks/aime24-x/README.md b/benchmarks/aime24-x/README.md
@@ -36,7 +36,7 @@ python benchmarks/aime24-x/prepare.py --prompt_language en
 ## Quickstart
 
 ```bash
-ng_run "+config_paths=[benchmarks/aime24-x/config.yaml,responses_api_models/vllm_model/configs/vllm_model.yaml]"
+ng_run "+config_paths=[benchmarks/aime24-x/config.yaml,responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml]"
 ```
 
 Then in another shell:

diff --git a/benchmarks/aime25-x/README.md b/benchmarks/aime25-x/README.md
@@ -36,7 +36,7 @@ python benchmarks/aime25-x/prepare.py --prompt_language en
 ## Quickstart
 
 ```bash
-ng_run "+config_paths=[benchmarks/aime25-x/config.yaml,responses_api_models/vllm_model/configs/vllm_model.yaml]"
+ng_run "+config_paths=[benchmarks/aime25-x/config.yaml,responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml]"
 ```
 
 Then in another shell:

diff --git a/benchmarks/aime26/README.md b/benchmarks/aime26/README.md
@@ -11,7 +11,7 @@ ng_prepare_benchmark "+config_paths=[benchmarks/aime26/config.yaml]"
 ## Run servers
 
 ```bash
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,benchmarks/aime26/config.yaml"
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,benchmarks/aime26/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 ```
 

diff --git a/benchmarks/answer-judge/README.md b/benchmarks/answer-judge/README.md
@@ -16,7 +16,7 @@ here is the same deterministic `Judgement: Yes/No` parsing used by Skills'
 ng_prepare_benchmark "+config_paths=[benchmarks/answer-judge/config.yaml]"
 
 # Running servers
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/answer-judge/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 

diff --git a/benchmarks/apex_shortlist/README.md b/benchmarks/apex_shortlist/README.md
@@ -35,7 +35,7 @@ Writes `data/apex_shortlist_benchmark.jsonl` with one row per problem:
 ## Running servers
 
 ```bash
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/apex_shortlist/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 ```

diff --git a/benchmarks/arena_hard/README.md b/benchmarks/arena_hard/README.md
@@ -31,7 +31,7 @@ to pick the standard judge prompt.
 ng_prepare_benchmark "+config_paths=[benchmarks/arena_hard/config.yaml]"
 
 # Running servers
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/arena_hard/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 

diff --git a/benchmarks/arena_hard_v2/README.md b/benchmarks/arena_hard_v2/README.md
@@ -34,7 +34,7 @@ repo, joins by `uid`, and emits one row per question with `question`,
 ng_prepare_benchmark "+config_paths=[benchmarks/arena_hard_v2/config.yaml]"
 
 # Running servers
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/arena_hard_v2/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 

diff --git a/benchmarks/asr_leaderboard/README.md b/benchmarks/asr_leaderboard/README.md
@@ -32,7 +32,7 @@ Downloads the 8 ESB subsets (~tens of GB of FLAC) and writes
 ## Running servers
 
 ```bash
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/asr_leaderboard/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 ```

diff --git a/benchmarks/bigcodebench/README.md b/benchmarks/bigcodebench/README.md
@@ -16,7 +16,7 @@ the `full` split (~1140 problems) is `bigcode/bigcodebench@v0.1.4`.
 ng_prepare_benchmark "+config_paths=[benchmarks/bigcodebench/config.yaml]"
 
 # Running servers
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/bigcodebench/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 

diff --git a/benchmarks/birdbench/README.md b/benchmarks/birdbench/README.md
@@ -22,7 +22,7 @@ and writes `data/birdbench_benchmark.jsonl`. Each row has
 ## Running servers
 
 ```bash
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/birdbench/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 ```

diff --git a/benchmarks/browsecomp/README.md b/benchmarks/browsecomp/README.md
@@ -8,7 +8,7 @@ browsecomp_benchmark_resources_server:
       exclude_domains_file_path: ???
 Qwen3-235B-A22B-Instruct-2507-FP8:
   responses_api_models:
-    vllm_model:
+    vllm_endpoint:
       base_url: ???
       api_key: ???
 ```
@@ -23,7 +23,7 @@ ng_prepare_benchmark "+config_paths=[$config_paths]"
 ```bash
 WANDB_PROJECT=
 EXPERIMENT_NAME=
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/browsecomp/config.yaml"
 ng_e2e_collect_rollouts \
     "+config_paths=[${config_paths}]" \

diff --git a/benchmarks/browsecomp/config.yaml b/benchmarks/browsecomp/config.yaml
@@ -1,7 +1,7 @@
-# Refer to the params in responses_api_models/local_vllm_model/configs/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8.yaml
+# Refer to the params in responses_api_models/vllm_server/configs/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8.yaml
 Qwen3-235B-A22B-Instruct-2507-FP8:
   responses_api_models:
-    vllm_model:
+    vllm_endpoint:
       entrypoint: app.py
       model: Qwen3-235B-A22B-Instruct-2507-FP8
       replace_developer_role_with_system: true
@@ -11,7 +11,7 @@ Qwen3-235B-A22B-Instruct-2507-FP8:
 policy_model_no_interleaved_reasoning:
   _copy: policy_model
   responses_api_models:
-    vllm_model:
+    vllm_endpoint:
       uses_interleaved_reasoning: false
 
 browsecomp_benchmark_resources_server:

diff --git a/benchmarks/finance_sec_search/README.md b/benchmarks/finance_sec_search/README.md
@@ -47,7 +47,7 @@ JSONL to `data/`.
 ## Running servers
 
 ```bash
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/finance_sec_search/config_no_web_search.yaml"
 ng_run "+config_paths=[$config_paths]"
 ```

diff --git a/benchmarks/flores200/README.md b/benchmarks/flores200/README.md
@@ -41,7 +41,7 @@ only advertised on multi-node SLURM deployments via NeMo-Skills'
 override and rely on corpus-BLEU only:
 
 ```bash
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/flores200/config.yaml"
 ng_run "+config_paths=[$config_paths]" \
     "++flores200_wmt_translation_resources_server.resources_servers.wmt_translation.compute_comet=false"
@@ -102,7 +102,7 @@ ns nemo_gym_rollouts \
     --server_nodes 2 \
     --server_args "--tensor-parallel-size 8 --data-parallel-size 1 --data-parallel-size-local 1 --data-parallel-backend ray --distributed-executor-backend ray --api-server-count 1 --reasoning-parser deepseek_r1 --trust-remote-code --dtype auto --enforce-eager" \
     --model "$MODEL" \
-    --config_paths "benchmarks/flores200/config.yaml,responses_api_models/vllm_model/configs/vllm_model.yaml" \
+    --config_paths "benchmarks/flores200/config.yaml,responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml" \
     --input_file benchmarks/flores200/data/flores200_devtest_benchmark.jsonl \
     --output_dir /workspace/flores200_smoke \
     --expname flores200_smoke \

diff --git a/benchmarks/frontierscience_olympiad/README.md b/benchmarks/frontierscience_olympiad/README.md
@@ -39,7 +39,7 @@ example, the original Skills configuration uses `o3-mini-2025-01-31` via
 ng_prepare_benchmark "+config_paths=[benchmarks/frontierscience_olympiad/config.yaml]"
 
 # Running servers
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/frontierscience_olympiad/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 

diff --git a/benchmarks/gdpval/README.md b/benchmarks/gdpval/README.md
@@ -19,7 +19,7 @@ ng_prepare_benchmark "+config_paths=[benchmarks/gdpval/config.yaml]"
 Each deliverable is scored 0–1 against the task rubric.
 
 ```bash
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/gdpval/config.yaml"
 ng_e2e_collect_rollouts \
     "+config_paths=[${config_paths}]" \

diff --git a/benchmarks/gdpval/config.yaml b/benchmarks/gdpval/config.yaml
@@ -3,7 +3,7 @@
 # Run:
 #   ng_prepare_benchmark "+config_paths=[benchmarks/gdpval/config.yaml]"
 #   ng_e2e_collect_rollouts \
-#     "+config_paths=[responses_api_models/vllm_model/configs/vllm_model.yaml,benchmarks/gdpval/config.yaml]" \
+#     "+config_paths=[responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,benchmarks/gdpval/config.yaml]" \
 #     ++split=benchmark \
 #     ++output_jsonl_fpath=results/gdpval.jsonl
 #

diff --git a/benchmarks/global-piqa/README.md b/benchmarks/global-piqa/README.md
@@ -19,7 +19,7 @@ server.
 ng_prepare_benchmark "+config_paths=[benchmarks/global-piqa/config.yaml]"
 
 # Running servers
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/global-piqa/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 

diff --git a/benchmarks/gpqa-x/README.md b/benchmarks/gpqa-x/README.md
@@ -37,7 +37,7 @@ python benchmarks/gpqa-x/prepare.py --prompt_language en
 ## Quickstart
 
 ```bash
-ng_run "+config_paths=[benchmarks/gpqa-x/config.yaml,responses_api_models/vllm_model/configs/vllm_model.yaml]"
+ng_run "+config_paths=[benchmarks/gpqa-x/config.yaml,responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml]"
 ```
 
 Then in another shell:

diff --git a/benchmarks/gpqa/README.md b/benchmarks/gpqa/README.md
@@ -16,10 +16,10 @@ This benchmark uses the `mcqa` resource server with the `mcqa_simple_agent`.
 ng_prepare_benchmark "+config_paths=[benchmarks/gpqa/config.yaml]"
 
 # Start servers
-ng_run "+config_paths=[benchmarks/gpqa/config.yaml,responses_api_models/vllm_model/configs/vllm_model.yaml]"
+ng_run "+config_paths=[benchmarks/gpqa/config.yaml,responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml]"
 
 # Collect rollouts
 ng_collect_rollouts \
-    "+config_paths=[benchmarks/gpqa/config.yaml,responses_api_models/vllm_model/configs/vllm_model.yaml]" \
+    "+config_paths=[benchmarks/gpqa/config.yaml,responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml]" \
     +output_jsonl_fpath=results/gpqa.jsonl
 ```
diff --git a/benchmarks/graphwalks/README.md b/benchmarks/graphwalks/README.md
@@ -47,7 +47,7 @@ python benchmarks/graphwalks/prepare.py \
 ## Start environment
 
 ```bash
-ng_run "+config_paths=[benchmarks/graphwalks/config.yaml,responses_api_models/vllm_model/configs/vllm_model.yaml]"
+ng_run "+config_paths=[benchmarks/graphwalks/config.yaml,responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml]"
 ```
 
 ## Collect rollouts

diff --git a/benchmarks/gsm8k/README.md b/benchmarks/gsm8k/README.md
@@ -16,7 +16,7 @@ the expected answer is integer-valued), then renames `problem` ->
 ng_prepare_benchmark "+config_paths=[benchmarks/gsm8k/config.yaml]"
 
 # Running servers
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/gsm8k/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 

diff --git a/benchmarks/hendrycks_math/README.md b/benchmarks/hendrycks_math/README.md
@@ -15,7 +15,7 @@ applies Skills' renames (`answer` -> `expected_answer`, `question` ->
 ng_prepare_benchmark "+config_paths=[benchmarks/hendrycks_math/config.yaml]"
 
 # Running servers
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/hendrycks_math/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 

diff --git a/benchmarks/hle/README.md b/benchmarks/hle/README.md
@@ -34,7 +34,7 @@ Downloads `cais/hle`, filters to text-only questions, and writes
 ## Running servers
 
 ```bash
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/hle/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 ```

diff --git a/benchmarks/hmmt_feb25/README.md b/benchmarks/hmmt_feb25/README.md
@@ -37,7 +37,7 @@ Start the benchmark's servers (inherits `math_with_judge` in symbolic-only mode
 plus a vLLM model server — adjust the model config to match your deployment):
 
 ```
-ng_run "+config_paths=[benchmarks/hmmt_feb25/config.yaml,responses_api_models/vllm_model/configs/vllm_model.yaml]"
+ng_run "+config_paths=[benchmarks/hmmt_feb25/config.yaml,responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml]"
 ```
 
 In a separate shell, collect rollouts against the full 30-problem set. `num_repeats`

diff --git a/benchmarks/hmmt_nov25/README.md b/benchmarks/hmmt_nov25/README.md
@@ -26,7 +26,7 @@ few-shots).
 Start vLLM with the `--reasoning-parser` that matches your model
 (e.g. `deepseek_r1` for models with a `<think>…</think>` convention;
 the parser name is declared in
-`responses_api_models/local_vllm_model/configs/nvidia/*.yaml`). Without
+`responses_api_models/vllm_server/configs/nvidia/*.yaml`). Without
 one, `math_with_judge` may extract intermediate expressions from
 truncated rollouts, and Skills' `parse_reasoning=True` default diverges
 on the same inputs.
@@ -38,7 +38,7 @@ on the same inputs.
 ng_prepare_benchmark "+config_paths=[benchmarks/hmmt_nov25/config.yaml]"
 
 # Running servers
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/hmmt_nov25/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 

diff --git a/benchmarks/hotpotqa_closedbook/README.md b/benchmarks/hotpotqa_closedbook/README.md
@@ -25,7 +25,7 @@ ng_prepare_benchmark "+config_paths=[benchmarks/hotpotqa_closedbook/config.yaml]
 ## Run servers
 
 ```bash
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,benchmarks/hotpotqa_closedbook/config.yaml"
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,benchmarks/hotpotqa_closedbook/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 ```
 

diff --git a/benchmarks/human_eval/README.md b/benchmarks/human_eval/README.md
@@ -23,7 +23,7 @@ holds the dataset definition + prompt + prepare script.
 ng_prepare_benchmark "+config_paths=[benchmarks/human_eval/config.yaml]"
 
 # Running servers
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/human_eval/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 

diff --git a/benchmarks/human_eval_infilling/README.md b/benchmarks/human_eval_infilling/README.md
@@ -40,7 +40,7 @@ holds only the dataset definition + prompt + prepare script.
 ng_prepare_benchmark "+config_paths=[benchmarks/human_eval_infilling/config.yaml]"
 
 # Running servers
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/human_eval_infilling/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 

diff --git a/benchmarks/ifbench/README.md b/benchmarks/ifbench/README.md
@@ -21,7 +21,7 @@ ng_prepare_benchmark "+config_paths=[benchmarks/ifbench/config.yaml]"
 
 ```bash
 ng_e2e_collect_rollouts \
-    "+config_paths=[benchmarks/ifbench/config.yaml,responses_api_models/vllm_model/configs/vllm_model.yaml]" \
+    "+config_paths=[benchmarks/ifbench/config.yaml,responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml]" \
     ++output_jsonl_fpath=results/benchmarks/ifbench.jsonl \
     ++overwrite_metrics_conflicts=true \
     ++split=benchmark \

diff --git a/benchmarks/ifeval/README.md b/benchmarks/ifeval/README.md
@@ -13,7 +13,7 @@ ng_prepare_benchmark "+config_paths=[benchmarks/ifeval/config.yaml]"
 ## Running servers
 
 ```bash
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/ifeval/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 ```

diff --git a/benchmarks/imo_answerbench/README.md b/benchmarks/imo_answerbench/README.md
@@ -13,7 +13,7 @@ ng_prepare_benchmark "+config_paths=[benchmarks/imo_answerbench/config.yaml]"
 ## Run servers
 
 ```bash
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,benchmarks/imo_answerbench/config.yaml"
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,benchmarks/imo_answerbench/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 ```
 

diff --git a/benchmarks/imo_gradingbench/README.md b/benchmarks/imo_gradingbench/README.md
@@ -36,7 +36,7 @@ details.
 ng_prepare_benchmark "+config_paths=[benchmarks/imo_gradingbench/config.yaml]"
 
 # Running servers
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/imo_gradingbench/config.yaml"
 ng_run "+config_paths=[$config_paths]"
 

diff --git a/benchmarks/imo_proofbench/README.md b/benchmarks/imo_proofbench/README.md
@@ -13,7 +13,7 @@ ng_prepare_benchmark "+config_paths=[benchmarks/imo_proofbench/config.yaml]"
 ## Run servers
 
 ```bash
-config_paths="responses_api_models/vllm_model/configs/vllm_model.yaml,\
+config_paths="responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml,\
 benchmarks/imo_proofbench/config.yaml"
 ng_run "+config_paths=[$config_paths]" \
     +judge_base_url=https://generativelanguage.googleapis.com/v1beta/openai \

diff --git a/benchmarks/ioi/README.md b/benchmarks/ioi/README.md
@@ -51,7 +51,7 @@ ng_prepare_data +config_paths=[benchmarks/ioi/config.yaml] \
   +output_dirpath=benchmarks/ioi/data \
   +mode=benchmark_preparation
 
-ng_run +config_paths=[benchmarks/ioi/config.yaml,responses_api_models/vllm_model/configs/vllm_model.yaml]
+ng_run +config_paths=[benchmarks/ioi/config.yaml,responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml]
 
 ng_collect_rollouts +agent_name=ioi_simple_agent \
   +input_jsonl_fpath=benchmarks/ioi/data/ioi24_benchmark.jsonl \

diff --git a/benchmarks/labbench2_vlm/README.md b/benchmarks/labbench2_vlm/README.md
@@ -19,7 +19,7 @@ the agent.
 The benchmark chains in `resources_servers/labbench2_vlm/configs/judge_model_openai.yaml`,
 which targets an OpenAI-compatible hosted endpoint. To use a different judge
 (e.g. a local vLLM), drop that chain and supply your own `responses_api_models`
-instance named `judge_model` — see `responses_api_models/vllm_model/configs/vllm_model.yaml`
+instance named `judge_model` — see `responses_api_models/vllm_endpoint/configs/vllm_endpoint.yaml`
 for the vLLM form.
 
 Credentials go in `env.yaml` at the **repository root** (the parser loads